Beruflich Dokumente
Kultur Dokumente
Numerical Approximation
Methods for Elliptic Boundary
Value Problems
Originally published in the German language by B.G. Teubner Verlag as “Olaf Steinbach:
Numerische Näherungsverfahren für elliptische Randwertprobleme. 1. Auflage (1st ed.)”.
c B.G. Teubner Verlag|GWV Fachverlage GmbH, Wiesbaden 2003
9 8 7 6 5 4 3 2 1
springer.com
Preface
Finite and boundary element methods belong to the most used numerical dis-
cretization methods for the approximate solution of elliptic boundary value
problems. Finite element methods (FEM) are based on a variational formu-
lation of the partial differential equation to be solved. The definition of a
conforming finite dimensional trial space requires an appropriate decomposi-
tion of the computational domain into finite elements. The advantage of using
finite element methods is their almost universal applicability, e.g. when con-
sidering nonlinear partial differential equations. Contrary, the use of boundary
element methods (BEM) requires the explicit knowledge of a fundamental so-
lution, which allows the transformation of the partial differential equation to a
boundary integral equation to be solved. The approximate solution then only
requires a decomposition of the boundary into boundary elements. Bound-
ary element methods are often used to solve partial differential equations
with (piecewise) constant coefficients, and to find solutions of boundary value
problems in exterior unbounded domains. In addition, direct boundary el-
ement methods provide a direct computation of the complete Cauchy data
which are the real target functions in many applications. In finite element
methods, the Cauchy data can be computed by using Lagrange multipliers
and by solving related saddle point problems. By combining both discretiza-
tion methods it is possible to profit from the advantages of both methods.
Although the aim of this book is to give a unified introduction into fi-
nite and boundary element methods, the main focus of the presentation is on
the numerical analysis of boundary integral equation methods. Therefore, we
only consider some linear model problems such as the potential equation, the
system of linear elasticity, the Stokes system, and the Helmholtz equation.
When considering the above mentioned elliptic boundary value problems it is
possible to describe and to analyze finite and boundary element methods in
a unified manner. After the description of the model problems, we introduce
the function spaces which are needed later. Then we discuss variational meth-
ods for the solution of operator equations with and without side conditions. In
particular, this also includes the formulation of saddle point problems by using
VI Preface
and proceeding to their practical realization see [117] where also numerous
examples are given.
Since the aim of this textbook is to give a unified introduction into finite
and boundary element methods, not all topics of interest can be discussed.
For a further reading we refer, e.g., to [51, 130] for hp finite element methods,
to [2, 8, 10, 153] for a posteriori error estimators and adaptive finite element
methods, and to [24, 69] for multigrid methods. In the case of boundary el-
ement methods we refer, e.g., to [17, 99, 100, 131, 146] for hp methods, to
[37, 38, 56, 128, 129] for a posteriori error estimators and adaptive methods,
and to [96, 111, 147] for multigrid and multilevel methods. Within this text-
book we also do not discuss the matter of numerical integration, for this we
refer to [55, 67, 87, 93, 115, 123, 132, 148] and the references given therein.
This textbook is based on lectures on the numerical solution of elliptic
partial differential equations which I taught at the University of Stuttgart, at
the Technical University of Chemnitz, at the Johannes Kepler University of
Linz, and at Graz University of Technology. Chapters 1–4, 8, 9, 11, and 13 can
be used for an introductory lecture on finite element methods, while chapters
1–8, 10, 12, and 13 are on the basics of the boundary element method. Chapter
14 gives an overview on fast boundary element methods. Besides the use as
a complementary textbook it is also recommended for self–study for students
and researchers, both in applied mathematics, in scientific computing, and in
computational engineering.
It is my great pleasure to thank W. L. Wendland for his encouragement
and support over the years. Many results of our joint work influenced this
book. Special thanks go to J. Breuer and G. Of, who read the original German
manuscript and made valuable comments and corrections. This text book was
originally published in a German edition [140]. Once again I would like to
thank J. Weiss and B. G. Teubner for the fruitful cooperation.
When preparing the English translation I got many responses, suggestions
and hints on the German edition. I would like to thank all who helped to
improve the book. In particular I thank G. Of, S. Engleder and D. Copeland
who read the English manuscript. Finally I thank Springer New York for the
cooperation and the patience when preparing this book.
2 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 The Spaces C k (Ω), C k,κ (Ω) and Lp (Ω) . . . . . . . . . . . . . . . . . . . . 19
2.2 Generalized Derivatives and Sobolev Spaces . . . . . . . . . . . . . . . . . 22
2.3 Properties of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Distributions and Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Sobolev Spaces on Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3 Variational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1 Operator Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Elliptic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Operators and Stability Conditions . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4 Operator Equations with Constraints . . . . . . . . . . . . . . . . . . . . . . 50
3.5 Mixed Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6 Coercive Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 Fundamental Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1 Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.2 Linear Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.3 Stokes Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.4 Helmholtz Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
1
Boundary Value Problems
The coefficient functions aji (x) are assumed to be sufficient smooth satisfying
aij (x) = aji (x) for all i, j = 1, . . . , d, x ∈ Ω. Partial differential operators
of the form (1.1) are used to model, for example, the static heat transfer,
electrostatic potentials, or ideal fluids.
For a classification [79] of scalar partial differential operators L we consider
the real eigenvalues λk (x) of the symmetric coefficient matrix
d
A(x) = (aij (x))i,j=1 , x ∈ Ω.
The partial differential operator L is called elliptic at x ∈ Ω iff λk (x) > 0 is
satisfied for all k = 1, . . . , d. If this condition is satisfied for all x ∈ Ω, then L
is elliptic in Ω. If there exists a uniform lower bound λ0 > 0 satisfying
λk (x) ≥ λ0 for k = 1, . . . , d and for all x ∈ Ω, (1.2)
the partial differential operator L is called uniformly elliptic in Ω.
2 1 Boundary Value Problems
The starting point for what follows is the well known theorem of Gauss
and Ostrogradski, i.e.
∂
f (x)dx = γ0int f (x)ni (x)dsx , i = 1, . . . , d,
∂xi
Ω Γ
where
γ0int f (x) := lim f (
x) for x ∈ Γ = ∂Ω (1.3)
x→x∈Γ
Ω
Example 1.1. For the special choice aij (x) = δij where δij is the Kronecker
delta with δij = 1 for i = j and δij = 0 for i = j, the partial differential
operator (1.1) is the Laplace operator
d
∂2
(Lu)(x) = −∆u(x) := − u(x) for x ∈ Rd . (1.9)
i=1
∂x2i
The associated conormal derivative (1.7) coincides with the normal derivative
∂
γ1int u(x) = u(x) := n(x) · ∇u(x) for x ∈ Γ.
∂nx
problem, while the boundary value problem (1.10) and (1.12) with Γ = ΓN is
called a Neumann boundary value problem. In the case Γ = ΓR the problem
(1.10) and (1.13) is said to be a Robin boundary value problem. In all other
cases we have to solve boundary value problems with boundary conditions of
mixed type. Note that one may also consider nonlinear Robin type boundary
conditions [119]
u ∈ C 2 (Ω) ∩ C 1 (Ω ∪ ΓN ∪ ΓR ) ∩ C(Ω ∪ ΓD )
where we have to assume that the given data are sufficiently smooth. For
results on the unique solvability of boundary value problems in the classical
sense we refer, for example, to [92].
In the case of a Neumann boundary value problem (1.10) and (1.12) addi-
tional considerations are needed to investigate the solvability of the boundary
value problem. Obviously, v1 (x) = 1 for x ∈ Ω is a solution of the homoge-
neous Neumann boundary value problem
When considering the Neumann boundary value problem (1.10) and (1.12),
and using the orthogonality (1.15) for the given data f and gN , we have to
assume the solvability condition
f (x)dx + gN (x)dsx = 0. (1.17)
Ω Γ
In (1.18), σij (u, x) are the components of the stress tensor which is linked to
the strain tensor eij (u, x) by Hooke’s law,
3
Eν E
σij (u, x) = δij ekk (u, x) + eij (u, x) (1.19)
(1 + ν)(1 − 2ν) 1+ν
k=1
for x ∈ Ω, i, j = 1, 2, 3, and with the Young modulus E > 0 and with the
Poisson ratio ν ∈ (0, 12 ). Moreover, when assuming small deformations the
linearized strain tensor is given by
1 ∂ ∂
eij (u, x) = uj (x) + ui (x) for x ∈ Ω, i, j = 1, 2, 3. (1.20)
2 ∂xi ∂xj
Multiplying the equilibrium equations (1.18) with a test function vi , integrat-
ing over Ω, and applying integration by parts, this gives for i = 1, 2, 3
3
∂
fi (x)vi (x)dx = − σij (u, x)vi (x)dx
j=1
∂xj
Ω Ω
3
3
∂
= σij (u, x) vi (x)dx − nj (x)σij (u, x)vi (x)dsx .
∂xj
Ω j=1 Γ j=1
Inserting the strain tensor (1.20) as well as Hooke’s law (1.19) into the equi-
librium equations (1.18) we obtain a system of partial differential equations
where the unknown function is the displacement field u. First we have
3
3
∂
ekk (u, x) = uk (x) =: div u(x)
∂xk
k=1 k=1
and therefore
Eν E ∂
σii (u, x) = div u(x) + ui (x),
(1 + ν)(1 − 2ν) 1 + ν ∂xi
E ∂ ∂
σij (u, x) = uj (x) + ui (x) for i = j.
2(1 + ν) ∂xi ∂xj
when exchanging the role of u and v. From the symmetry of the bilinear form
a(·, ·) we then obtain Betti’s second formula
3
∂
− σij (u, x)vi (x)dx + γ0int v(x) γ1int u(x)dsx (1.30)
i,j=1
∂xj
Ω Γ
3
∂
= − σij (v, x)ui (x)dx + γ0int u(x) γ1int v(x)dsx .
i,j=1
∂x j
Ω Γ
Inserting the rigid body motions v k ∈ R into Betti’s second formula (1.30)
this gives the orthogonality
3
∂
− σij (u, x)vk,i (x)dx + γ0int v k (x) γ1int u(x)dsx = 0
i,j=1
∂xj
Ω Γ
for all v k ∈ R. Hence, for the solvability of the Neumann boundary value
problem
Note that the solution of the Neumann boundary value problem is only unique
up to the rigid body motions (1.29). A unique solution can be defined when
considering either nodal or scaling conditions in addition.
The second order system (1.25) of linear elasticity can also be written as
a scalar partial differential equation of fourth order. By setting
λ+µ
u(x) := ∆w(x) − grad div w(x) for x ∈ Ω (1.32)
λ + 2µ
1.2 Linear Elasticity 9
from the equilibrium equations (1.25) we obtain the scalar Bi–Laplace equa-
tion
−µ∆2 w(x) = f (x) for x ∈ Ω. (1.33)
When considering a homogeneous partial differential equation with f ≡ 0
the solution of (1.25) can be described by setting w2 ≡ w3 ≡ 0 where the
remaining component w1 = ψ is the solution of the Bi–Laplace equation
Then,
λ + µ ∂2
u1 (x) := ∆ψ(x) − ψ(x),
λ + 2µ ∂x21
λ+µ ∂2
u2 (x) := − ψ(x),
λ + 2µ ∂x1 ∂x2
λ+µ ∂2
u3 (x) := − ψ(x)
λ + 2µ ∂x1 ∂x3
is a solution of the homogeneous system (1.25). The function ψ is known as
Airy’s stress function.
To describe problems of linear elasticity in two space dimensions one may con-
sider two different approaches. In plain stress we assume that the stress tensor
depends on two space coordinates (x1 , x2 ) only, and that the x3 –coordinates
of the stress tensor disappear:
and
ν
e33 (u, x) = − [e11 (u, x) + e22 (u, x)]. (1.35)
1−ν
The resulting stress–strain relation reads
2
Eν E
σij (u, x) = δij ekk (u, x) + eij (u, x)
(1 + ν)(1 − ν) 1+ν
k=1
u3 (x1 , x2 , x3 ) = u3 (x3 ) .
and therefore
ν ∂ ∂
u3 (x) = − u1 (x) + u2 (x) x3 .
1 − ν ∂x1 ∂x2
= Eν E
λ , =
µ
(1 + ν)(1 − ν) 2(1 + ν)
− +µ
µ∆u(x) − (λ )grad div u(x) = f (x) for x ∈ Ω ⊂ R2 .
In plain strain we assume that all components eij (u, x) of the strain tensor
depend only on the space coordinates (x1 , x2 ) and that the x3 –coordinates
vanish:
and
1.2 Linear Elasticity 11
Eν
σ33 (u, x) = [e11 (u, x) + e22 (u, x)] .
(1 + ν)(1 − 2ν)
With
E
σ11 (u, x) + σ22 (u, x) = [e11 (u, x) + e22 (u, x)]
(1 + ν)(1 − 2ν)
we finally get
σ33 (u, x) = ν [σ11 (u, x) + σ22 (u, x)] .
In both cases of two–dimensional plane stress and plane strain linear elasticity
models the rigid body motions are given as
1 0 −x2
R = span , , . (1.36)
0 1 x1
∂
γ1int u(x) = [λ div u(x)]n(x) + 2µ u(x) + µ n(x) × curl v(x), x ∈ Γ.
∂nx
12 1 Boundary Value Problems
which coincides with the Stokes system which plays an important role in fluid
mechanics.
When considering the Stokes system [91] we have to find a velocity field u and
the pressure p satisfying the system of partial differential equations
Therefore, the given Dirichlet data g(x) have to satisfy the solvability condi-
tion (1.39.) Moreover, the pressure p is only unique up some additive constant.
When multiplying the components of the first partial differential equa-
tion in (1.38) with some test function vi , integrating over Ω, and applying
integration by parts this gives
∂
fi (x)vi (x)dx = −µ ∆ui (x)vi (x)dx + p(x)vi (x)dx (1.40)
∂xi
Ω Ω Ω
∂
= −µ ∆ui (x)vi (x)dx − p(x) vi (x)dx + p(x)ni (x)vi (x)dsx .
∂xi
Ω Ω Γ
1.3 Stokes System 13
Moreover we have
∂ ∂ ∂
[eij (u, x)vi (x)] = vi (x) eij (u, x) + eij (u, x) vi (x),
∂xj ∂xj ∂xj
as well as
d
∂ d
eij (u, x) vi (x) = eij (u, x)eij (v, x).
i,j=1
∂xj i,j=1
3
∂ ∂ ∂ ∂
e1j (u, x) = e11 (u, x) + e12 (u, x) + e13 (u, x)
j=1
∂xj ∂x1 ∂x2 ∂x3
∂2 1 ∂ ∂ ∂
= u 1 (x) + u 2 (x) + u 1 (x)
∂x21 2 ∂x2 ∂x1 ∂x2
1 ∂ ∂ ∂
+ u3 (x) + u1 (x)
2 ∂x3 ∂x1 ∂x3
1 1 ∂ ∂ ∂ ∂
= ∆u1 (x) + u1 (x) + u2 (x) + u3 (x)
2 2 ∂x1 ∂x1 ∂x2 ∂x3
1 1 ∂
= ∆u1 (x) + div u(x).
2 2 ∂x1
Corresponding results hold for i = 2, 3 and d = 2, respectively. Then we obtain
d
∂ d
∂ d
∂
[eij (u, x)vi (x)] = eij (u, x)vi (x) + eij (u, x) vi (x)
i,j=1
∂x j i,j=1
∂x j i,j=1
∂x j
d
1 d
∂
= ∆ui (x) + div u(x) vi (x) + eij (u, x)eij (v, x).
2 i=1 ∂xi i,j=1
d
∂
+ vi (x) div u(x).
i=1
∂xi
Taking the sum of (1.40) for i = 1, . . . , d, substituting the above results, and
applying integration by parts this gives
14 1 Boundary Value Problems
d
d
∂
v(x) f (x)dx = −µ vi (x)∆ui (x)dx + vi (x) p(x) dx
∂xi
Ω Ω i=1 Ω i=1
d
d
∂
= 2µ eij (u, x)eij (v, x)dx − 2µ [eij (u, x)vi (x)] dx
∂xj
Ω i,j=1 Ω i,j=1
d
d
∂ ∂
+µ vi (x) div u(x)dx + vi (x) p(x) dx
i=1
∂x i ∂xi
Ω Ω i=1
d
d
= 2µ eij (u, x)eij (v, x)dx − 2µ nj (x)eij (u, x)γ0int vi (x)dsx
Ω i,j=1 Γ i,j=1
−µ div u(x) div v(x) dx + µ div u(x)n(x) γ0int v(x)dsx
Ω Γ
1 ∂2
U (x, t) = ∆U (x, t) for x ∈ Rd (d = 2, 3)
c2 ∂t2
describes the wave propagation in a homogeneous, isotropic and friction–free
medium having the constant speed of sound c. Examples are acoustic scatter-
ing and sound radiation problems.
For time harmonic acoustic waves
U (x, t) = Re u(x)e−iωt
and by using the symmetry of the bilinear form a(·, ·) we conclude Green’s
second formula
2
[−∆u(x) − k u(x)]v(x)dx + γ1int u(x)γ0int v(x)dsx (1.46)
Ω Γ
= [−∆v(x) − k 2 v(x)]u(x)dx + γ1int v(x)γ0int u(x)dsx .
Ω Γ
For any solution u of the Helmholtz equation (1.44) we find from (1.45) by
setting v = u
|∇u(x)|2 dx − k 2 |u(x)|2 dx = γ1int u(x)γ0int u(x)dsx . (1.47)
Ω Ω Γ
For x0 ∈ Ω let BR (x0 ) be a ball with center x0 and radius R such that
Ω ⊂ BR (x0 ) is satisfied. Then, ΩR = BR (x0 )\Ω is a bounded domain for
which we can write (1.47) as
2 2
|∇u(x)| dx − k |u(x)|2 dx
ΩR ΩR
= − γ1ext u(x)γ0ext u(x)dsx + γ1int u(x)γ0int u(x)dsx
Γ ∂BR (x0 )
taking into account the opposite direction of the normal vector on Γ . This
clearly implies
Im γ1int u(x)γ0int u(x)dsx = Im γ1ext u(x)γ0ext u(x)dsx = O(1).
∂BR (x0 ) Γ
On the other hand, from the Sommerfeld radiation condition (1.48) we also
conclude the weaker condition due to Rellich,
2
int
lim γ1 u(x) − ikγ0int u(x) dsx = 0. (1.49)
R→∞
∂BR (x0 )
1.5 Exercises 17
and therefore
2k Im γ1ext u(x)γ0ext u(x)dsx
Γ
⎡ ⎤
2 2
⎢ int int ⎥
= lim ⎣ γ1 u(x) dsx + k 2 γ0 u(x) dsx ⎦ ≥ 0
R→∞
∂BR (x0 ) ∂BR (x0 )
implying 2
int
lim γ0 u(x) dsx = O(1)
R→∞
∂BR (x0 )
as well as
1
|u(x)| = O as |x| → ∞. (1.50)
|x|
1.5 Exercises
1.1 For x ∈ R2 we consider polar coordinates
(r, ϕ).
u(x1 , x2 ) = u(x1 (r, ϕ), x2 (r, ϕ)) = u
18 1 Boundary Value Problems
∇(r,ϕ) u
(r, ϕ) = J ∇x u(x).
∂2 ∂2
∆u(x) = u(x) + u(x)
∂x21 ∂x22
(r, ϕ) = rα sin(αϕ)
u(x) = u
∂2 ∂2 ∂2
∆u(x) = 2 u(x) + 2 u(x) + u(x) for x ∈ R3
∂x1 ∂x2 ∂x23
Determine the constant α ∈ R such that the solution of the Navier system
Γ := ∂Ω = Ω ∩ (Rd \Ω) .
If γ(·) is Lipschitz,
|γ(
x) − γ(
y )| ≤ L |
x − y| for all x
, y ∈ Rd−1
Two elements u, v ∈ Lp (Ω) are identified with each other if they are different
only on a set K of zero measure µ(K) = 0. In what follows we always consider
one represent u ∈ Lp (Ω). In addition, L∞ (Ω) is the space of functions u which
are measurable and bounded almost everywhere with the norm
The spaces Lp (Ω) are Banach spaces with respect to the norm · Lp (Ω) .
There holds the Minkowski inequality
1 1
+ = 1,
p q
we further have Hölder’s inequality
|u(x)v(x)|dx ≤ uLp (Ω) vLq (Ω) . (2.2)
Ω
22 2 Function Spaces
we obtain
|u, vΩ | 1 1
vLq (Ω) = sup for 1 ≤ p < ∞, + = 1.
0=u∈Lp (Ω) uLp (Ω) p q
and with
u, uL2 (Ω) = u2L2 (Ω) for all u ∈ L2 (Ω)
we conclude that L2 (Ω) is a Hilbert space.
Note that all integrals may be defined even for non–smooth functions. This
motivates the following definition of a generalized derivative.
∂
Again we denote the generalized derivative by u(x) := v(x).
∂xi
The recursive application of (2.4) enables us to define a generalized partial
derivative Dα u(x) ∈ Lloc
1 (Ω) by
|α|
α
[D u(x)]ϕ(x)dx = (−1) u(x)Dα ϕ(x)dx for all ϕ ∈ C0∞ (Ω). (2.5)
Ω Ω
Example 2.4. Let u(x) = |x| for x ∈ Ω = (−1, 1). For an arbitrary ϕ ∈ C0∞ (Ω)
we have
1 0 1
∂ ∂ ∂
u(x) ϕ(x)dx = − x ϕ(x)dx + x ϕ(x)dx
∂x ∂x ∂x
−1 −1 0
0 1
0 1
= − [x ϕ(x)]−1 + ϕ(x)dx + [x ϕ(x)]0 − ϕ(x)dx
−1 0
0 1 1
= ϕ(x)dx − ϕ(x)dx = − sign(x) ϕ(x)dx
−1 0 −1
with
1 for x > 0,
sign(x) :=
−1 for x < 0.
The generalized derivative of u(x) = |x| is therefore given by
∂
u(x) = sign(x) ∈ Lloc
1 (Ω).
∂x
To compute the second derivative of u(x) = |x| we obtain
24 2 Function Spaces
1 0 1
∂ ∂ ∂
sign(x) ϕ(x)dx = − ϕ(x)dx + ϕ(x)dx = −2ϕ(0) .
∂x ∂x ∂x
−1 −1 0
1
v(x)ϕ(x)dx = 2ϕ(0)
−1
for all ϕ ∈ C0∞ (Ω). Later we will find the generalized derivative of sign(x) in
the distributional sense.
By taking the closure of C ∞ (Ω) with respect to the norm · Wpk (Ω) we define
the Sobolev space
·
W k (Ω)
Wpk (Ω) := C ∞ (Ω) p . (2.7)
In particular, for any u ∈ Wpk (Ω) there exists a sequence {ϕj }j∈N ⊂ C ∞ (Ω)
such that
lim u − ϕj Wpk (Ω) = 0.
j→∞
Correspondingly, the closure of C0∞ (Ω) with respect to · Wpk (Ω) defines the
Sobolev space
◦
·
W k (Ω)
W p (Ω) := C0∞ (Ω)
k p . (2.8)
The definition of Sobolev norms · Wpk (Ω) and therefore of the Sobolev spaces
(2.7) and (2.8) can be extended for any arbitrary s ∈ R. We first consider
0 < s ∈ R with s = k + κ and k ∈ N0 , κ ∈ (0, 1). Then,
! "1/p
uWps (Ω) := upW k (Ω) + |u|pW s (Ω)
p p
for s = k ∈ N0 and
u, vW2s (Ω) := u, vW2k (Ω) (2.9)
◦
Correspondingly, W ps (Ω) is the dual space of Wq−s (Ω).
For a proof of Theorem 2.5 see, for example, [31, Theorem 1.4.6], [103,
Theorem 3.26].
The norm (2.6) of the Sobolev space W21 (Ω) is
! "1/2
vW21 (Ω) = v2L2 (Ω) + ∇v2L2 (Ω)
where
|v|W21 (Ω) = ∇vL2 (Ω)
is a semi–norm. Applying the following theorem we may deduce equivalent
norms in W21 (Ω).
26 2 Function Spaces
The proof of the opposite direction is indirect. Assume that there is no con-
stant c0 > 0 such that
Then there would exist a sequence {vn }n∈N ⊂ W21 (Ω) with
vn W21 (Ω)
n ≤ for n ∈ N.
vn W21 (Ω),f
we therefore have
vn W21 (Ω),f 1
v̄n W21 (Ω) = 1, v̄n W21 (Ω),f = ≤ →0 as n → ∞.
vn W21 (Ω) n
Since the sequence {v̄n }n∈N is bounded in W21 (Ω) and since the imbedding
W21 (Ω) → L2 (Ω) is compact [160], there exists a subsequence {v̄n }n ∈N ⊂
{v̄n }n∈N which converges in L2 (Ω). In particular, v̄ := lim
v̄n ∈ L2 (Ω).
n →∞
From
lim ∇v̄n L2 (Ω) = 0
n →∞
0 ≤ |f (v̄)| = lim
|f (v̄n )| = 0
n →∞
and ⎧⎡ ⎫1/2
⎤2
⎪
⎨ ⎪
⎬
vW21 (Ω),Γ := ⎣ ⎦ 2
v(x) dsx + ∇vL2 (Ω) .
⎪
⎩ ⎪
⎭
Γ
◦
We therefore obtain ∇ · L2 (Ω) to be an equivalent norm in W12 (Ω)
Using the equivalent norm (2.11) as given in Example 2.7 the Poincaré
inequality
⎧⎡ ⎤2 ⎫
⎪
⎨ ⎪
⎬
2
|v(x)| dx ≤ cP ⎣ ⎦
v(x) dx + |∇v(x)| dx2
(2.12)
⎪
⎩ ⎪
⎭
Ω Ω Ω
f (q) = 0
where the constant c(cp ) depends only on the constant cp of the Poincaré
inequality (2.12).
28 2 Function Spaces
Proof. We give only the proof for the case k = 1, i.e. P1 (Ω) is the space of
linear functions defined on Ω. For v ∈ W22 (Ω) and q ∈ P1 (Ω) we have, due to
the assumptions,
Moreover,
For the second term we apply the Poincaré inequality (2.12) once again to
obtain
d 2
∂
2
∇(v + q)L2 (Ω) =
∂xi [v(x) + q(x)] dx
i=1 Ω
⎧⎡ ⎤2 ⎫
d ⎪⎨ ∂ d 2
2 ⎪⎬
⎣ ∂
≤ cP [v(x) + q(x)]dx⎦ + [v(x) + q(x)] dx
⎪ ∂xi ∂xi ∂xj ⎪
i=1 ⎩ Ω j=1 Ω
⎭
⎡ ⎤2
d
⎣ ∂
= cP [v(x) + q(x)]dx⎦ + cP |v|2W 2 (Ω)
i=1
∂xi 2
and hence
⎡ ⎤2
v + q2W 2 (Ω) ≤ cP ⎣ [v(x) + q(x)]dx⎦
2
Ω
⎡ ⎤2
d
⎣ ∂
+(1 + cP )cP [v(x) + q(x)]dx⎦ + [1 + (1 + cP )cP ] |v|2W 2 (Ω) .
i=1
∂xi 2
The assertion is proved if we can choose q ∈ P1 (Ω) so that the first two terms
are zero, i.e.
∂
[v(x) + q(x)]dx = 0, [v(x) + q(x)]dx = 0 for i = 1, . . . , d.
∂xi
Ω Ω
2.4 Distributions and Sobolev Spaces 29
With
d
q(x) = a0 + ai xi
i=1
we get
1 ∂
ai = − v(x)dx for i = 1, . . . , d
|Ω| ∂xi
Ω
# $
and therefore
1
d
a0 = − v(x) + ai xi dx.
|Ω| i=1
Ω
For u ∈ Lloc
1 (Ω) we define the distribution
Tu (ϕ) := u(x)ϕ(x)dx for ϕ ∈ D(Ω). (2.13)
Ω
Distributions of the type (2.13) are called regular. Local integrable func-
tions u ∈ Lloc
1 (Ω) can be identified with a subset of D (Ω). Hence, instead
of Tu ∈ D (Ω) we simply write u ∈ D (Ω). Nonregular distributions are called
singular. For example, the Dirac distribution for x0 ∈ Ω,
1 1
∂ ∂
sign(x) ϕ(x)dx = −2ϕ(0) = − v(x) ϕ(x)dx for all ϕ ∈ D(Ω).
∂x ∂x
−1 −1
In particular, the function ϕ and all of their derivatives decreases faster than
any polynomial.
2
Example 2.12. For the function ϕ(x) := e−|x| we have ϕ ∈ S(Rd ), but
ϕ ∈ D(Ω) = C0∞ (Rd ).
The mapping F : S(Rd ) → S(Rd ) is invertible and the inverse Fourier trans-
form is given by
−1 −d
%
(F ϕ)(x) = (2π) 2 %
eix,ξ
ϕ(ξ)dξ for x ∈ Rd . (2.15)
Rd
% ∈ D(Rd ).
In general, ϕ ∈ D(Rd ) does not imply ϕ
2.4 Distributions and Sobolev Spaces 31
as well as
ξ α (Fϕ)(ξ) = (−i)|α| F(Dα ϕ)(ξ). (2.17)
∞ 2π
1
%(|ξ|, ψ + ψ0 ) =
u e−ir|ξ| cos(φ−ψ−ψ0 ) u(r)rdφdr
2π
0 0
∞ 2π−ψ
0
1
= e−ir|ξ| cos(φ−ψ) u(r)rdφdr.
2π
0 −ψ0
By using
0 2π
−ir|ξ| cos(φ−ψ)
e dφ = e−ir|ξ| cos(φ−ψ) dφ
−ψ0 2π−ψ0
we then obtain
%(|ξ|, ψ) = u
u %(|ξ|, ψ + ψ0 ) for all ψ0 ∈ [0, 2π)
%(ξ) = u
and therefore the assertion u %(|ξ|).
32 2 Function Spaces
to obtain
%(|ξ|, ψ, ϑ) =
u
∞ 2ππ
1
= e−ir|ξ|[cos(φ−ψ) sin θ sin ϑ+cos θ cos ϑ] u(r)r2 sin θdθdφdr.
(2π)3/2
0 0 0
%(|ξ|, ψ + ψ0 , ϑ) = u
u %(|ξ|, ψ, ϑ) for all ψ0 ∈ [0, 2π).
T%(ϕ) := T (ϕ)
% for ϕ ∈ S(Rd ).
The mapping F : S (Rd ) → S (Rd ) is invertible and the inverse Fourier trans-
form is given by
The rules (2.16) and (2.17) remain valid for distributions T ∈ S (Rd ).
For s ∈ R and u ∈ S(Rd ) we define the Bessel potential operator
−d/2
J u(x) := (2π)
s
(1 + |ξ|2 )s/2 u
%(ξ)eix,ξ
dξ, x ∈ Rd ,
Rd
The Sobolev space H s (Rd ) is the space of all distributions v ∈ S (Rd ) with
J s v ∈ L2 (Rd ) where the associated inner product
The connection with the Sobolev spaces W2s (Rd ) can be seen from the follow-
ing theorem, see for example [103, 160].
Theorem 2.14. For all s ∈ R there holds
s (Ω) := C ∞ (Ω)
·
H s (Rd ) ,
H H0s (Ω) := C0∞ (Ω)
·
H s (Ω)
0
which will coincide for almost all s ∈ R+ , see for example [103, Theorem 3.33].
Theorem 2.15. Let Ω ⊂ Rd be a Lipschitz domain. For s ≥ 0 we have
s (Ω) ⊂ H s (Ω).
H 0
In particular,
s (Ω) = H s (Ω) 1 3 5
H 0 for s ∈ , , ,... .
2 2 2
Moreover,
s (Ω) = [H −s (Ω)] ,
H −s (Ω)]
H s (Ω) = [H for all s ∈ R.
34 2 Function Spaces
The equivalence of Sobolev spaces W2s (Ω) and H s (Ω) holds only when
certain assumptions on Ω are satisfied. Sufficient for the norm equivalence is
the existence of a linear bounded extension operator
EΩ : W2s (Ω) → W2s (Rd ).
This condition is ensured for a bounded domain Ω ⊂ Rd , if a uniform cone
condition is satisfied, see for example [161, Theorem 5.4].
Theorem 2.16. For a Lipschitz domain Ω ⊂ Rd we have
W2s (Ω) = H s (Ω) for all s > 0.
For the analysis of the Stokes system we need to have some mapping
properties of the gradient ∇.
Theorem 2.17. [53, Theorem 3.2, p. 111] Let the Lipschitz domain Ω ⊂ Rd
be bounded and connected. Then there holds
qL2 (Ω) ≤ c1 qH −1 (Ω) + ∇q[H −1 (Ω)]d for all q ∈ L2 (Ω)
as well as
qL2 (Ω) ≤ c2 ∇q[H −1 (Ω)]d for all q ∈ L2 (Ω) with q(x)dx = 0. (2.18)
Ω
*
J
Γ = Γi , Γi := x ∈ Rd : x = χi (ξ) for ξ ∈ Ti ⊂ Rd−1 . (2.19)
i=1
With respect to (2.19) we also consider a partition of unity, {ϕi }pi=1 , of non–
negative cut off functions ϕi ∈ C0∞ (Rd ) satisfying
J
ϕi (x) = 1 for x ∈ Γ, ϕi (x) = 0 for x ∈ Γ \Γi .
i=1
J
J
v(x) = ϕi (x)v(x) = vi (x) for x ∈ Γ
i=1 i=1
The functions vi are defined with respect to the parameter domains Ti ⊂ Rd−1
for which we can introduce appropriate Sobolev spaces. Taking into account
the chain rule we have to ensure the existence of all corresponding derivatives
of the local parametrization χi (ξ). For the definition of derivatives of order
|s| ≤ k we therefore have to assume χi ∈ C k−1,1 (Ti ). In particular for a
Lipschitz domain with a local parametrization χi ∈ C 0,1 (Ti ) we can only
introduce Sobolev spaces H s (Ti ) for |s| ≤ 1.
In general we can define the Sobolev norm
J +1/2
vHχs (Γ ) := vi 2H s (Ti )
(2.20)
i=1
and
J
v2L2 (Γ ) = [v(x)]2 dsx = ϕi (x)[v(x)]2 dsx .
Γ i=1 Γ
i
From this the assertion follows, where the constants depend on both the chosen
parametrization (2.19) and on the particular definition of the cut off functions
ϕi .
For s ∈ (0, 1) we find in the same way that the Sobolev–Slobodeckii norm
⎧ ⎫1/2
⎨ 2 ⎬
[v(x) − v(y)]
vH s (Γ ) := v2L2 (Γ ) + ds ds
x y
⎩ |x − y|d−1+2s ⎭
Γ Γ
H s (Γ ) := [H −s (Γ )] ,
*
J
Γ = Γ i, Γi ∩ Γj = ∅ for i = j.
i=1
,
J
s
Hpw (Γ ) := s (Γj )
H (2.21)
j=1
wH s (Γ ) ≤ wHpw
s (Γ ) .
38 2 Function Spaces
|w, vΓ | J
|w, vΓj |
wH s (Γ ) = sup ≤ sup
−s
0=v∈H (Γ ) v H −s (Γ ) −s
0=v∈H (Γ ) j=1 vH −s (Γ )
J
|w|Γj , v|Γj Γj |
≤ sup
0=v∈H −s (Γ ) j=1 v|Γj H −s (Γj )
J
|w|Γj , vj Γj |
≤ sup = wHpw
s (Γ ) .
j=1 0=vj ∈H −s (Γ j)
vj H −s (Γj )
Remark 2.23. The interpolation theorem (Theorem 2.18) holds also for appro-
priate Sobolev spaces H s (Γ ).
If
f (q) = 0
is satisfied for all q ∈ Pk (Γ ) then we also have
2.6 Exercises
1 1
2
[u(x)] dx ≤ c [u (x)]2 dx
0 0
1 1
[u(x) − u(y)]2
dx dy < ∞
|x − y|1+2s
0 0
is finite.
3
Variational Methods
. space with the inner product ·, ·X and with the induced
Let X be a Hilbert
norm · X = ·, ·X . Let X be the dual space of X with respect to the
duality pairing ·, ·. Then it holds that
|f, v|
f X = sup for all f ∈ X . (3.1)
0=v∈X vX
AvX ≤ cA
2 vX for all v ∈ X. (3.2)
Au = f . (3.3)
|Au − f, v|
Au − f X = sup = 0
0=v∈X vX
a(·, ·) : X × X → R. (3.5)
2 uX vX
|a(u, v)| ≤ cA for all u, v ∈ X.
AuX ≤ cA
2 uX for all u ∈ X.
d
F (u + tv)|t=0 = 0 for all v ∈ X.
dt
From this we obtain
and therefore the equivalence of both the variational and the minimization
problem.
To investigate the unique solvability of the operator equation (3.3) we now
consider a fixed point iteration. For this we need to formulate the following
Riesz representation theorem.
uX = f X . (3.7)
44 3 Variational Methods
we then obtain
0 ≤ uk − u 2X = 2 uk 2X + 2 u 2X − uk + u 2X
1 2 1 2
=4 uk X − f, uk + 4 u X − f, u
2 2
+4 f, uk + u − uk + u 2X
1
= 4 F (uk ) + 4 F (u ) − 8 F (uk + u )
2
≤ 4 F (uk ) + 4 F (u ) − 8α → 0 as k, → ∞.
Moreover,
1
|F (uk ) − F (u)| ≤ |uk , uk X − u, uX | + |f, uk − u|
2
1
= |uk , uk − uX + u, uk − uX | + |f, uk − u|
2
1 1
≤ uk X + uX + f X uk − uX ,
2 2
and hence
F (u) = lim F (uk ) = α.
k→∞
u, vX = f, v for all v ∈ X.
u − u
, vX = 0 for all v ∈ X.
Choosing v = u − u
we now obtain
2X = 0
u − u
Moreover,
Jf X = f X . (3.11)
46 3 Variational Methods
To ensure the unique solvability of the operator equation (3.3) and of the vari-
ational problem (3.4) we need to have a further assumption for the operator
A and for the bilinear form a(·, ·), respectively. The operator A : X → X is
called X–elliptic if
2
Av, v ≥ cA
1 vX for all v ∈ X (3.12)
u = u − J(Au − f ) = T u + Jf
and therefore
T v2X = (I − JA)v2X
A 2
Hence we obtain that for ∈ (0, 2cA
1 /(c2 ) ) the operator T is a contraction
in X, and the unique solvability of (3.3) follows from Banach’s fixed point
theorem [163]. Let u ∈ X be the unique solution of the operator equation
(3.3). Then,
2
1 uX ≤ Au, u = f, u ≤ f X uX ,
cA
Applying Theorem 3.4 this gives the inverse operator A−1 : X → X and
we obtain
1
A−1 f X ≤ A f X for all f ∈ X . (3.13)
c1
From the boundedness of the self–adjoint and invertible operator A we also
conclude an ellipticity estimate for the inverse operator A−1 .
AvX ≤ cA
2 vX for all v ∈ X.
Then,
1
A−1 f, f ≥ f 2X for all f ∈ X .
cA
2
BvX = JAvX = AvX ≤ cA
2 vX for all v ∈ X.
Since
Bu, vX = JAu, vX = Au, v = u, Av = u, BvX
holds for all u, v ∈ X the operator B is self–adjoint satisfying the ellipticity
estimate
2
Bv, vX = Av, v ≥ cA1 vX for all v ∈ X.
Hence there exists a self–adjoint and invertible operator B 1/2 satisfying
B = B 1/2 B 1/2 , see, e.g., [118]. In addition we define B −1/2 := (B 1/2 )−1 .
Then we obtain
and further /
B 1/2 vX ≤ 2 vX
cA for all v ∈ X.
For an arbitrary f ∈ X we then conclude
and therefore
f 2X ≤ cA
2 B
−1/2
Jf 2X = cA
2 B
−1
Jf, Jf X = cA
2 A
−1
f, f .
48 3 Variational Methods
Proof. Here we only prove that iii. follows from i., see also [163]. From the
definition of the polar space with respect to B we find that
and therefore
ImX B ⊂ (ker B )0 .
Let g ∈ (ker B )0 with g ∈ ImX B. Applying the separation theorem for closed
convex sets there exists a q̄ ∈ Π and a constant α ∈ R such that
g, q̄ > α > f, q̄ for all f ∈ ImX (B) ⊂ Π .
For any f ∈ ImX (B) there exists at least one u ∈ X with f = Bu. Hence,
and therefore q̄ ∈ ker B . On the other hand, for g ∈ (ker B)0 we have
and therefore g, q̄ = 0 which is a contradiction to g, q̄ > α > 0.
The solvability condition (3.17) is equivalent to
If the equivalent solvability conditions (3.17) and (3.18) are satisfied, then
there exists at least one solution u ∈ X satisfying Bu = g. When the null
space ker B is non–trivial, we can add an arbitrary u0 ∈ ker B, in particular,
u + u0 is still a solution of B(u + u0 ) = g. In this case, the solution is not
unique in general. Instead we consider only solutions u ∈ (ker B)⊥ . To ensure
unique solvability in this case, we have to formulate additional assumptions.
50 3 Variational Methods
For a given g ∈ ImX (B) there exists a unique solution u ∈ (ker B)⊥ of the
operator equation Bu = g satisfying
1
uX ≤ gΠ .
cS
Proof. Since we assume g ∈ ImX B there exists at least one solution
u ∈ (ker B)⊥ of the operator equation Bu = g satisfying
Then,
B(u − ū), q = 0 for all q ∈ Π.
Obviously, u − ū ∈ (ker B)⊥ . From the stability condition (3.19) we then
conclude
B(u − ū), q
0 ≤ cS u − ūX ≤ sup = 0
0=q∈Π qΠ
and therefore uniqueness, u = ū. Applying (3.19) for the solution u this gives
Bu, q g, q
cS uX ≤ sup = sup ≤ gΠ .
0=q∈Π qΠ 0=q∈Π qΠ
Vg := {v ∈ X : Bv = g} .
The unique solvability of (3.20) now follows from the following result.
Au0 = f − Aug
we obtain
A(u0 + ug − û0 − ûg ) = 0.
Obviously, u0 − û0 + (ug − ûg ) ∈ V0 , and from the V0 –ellipticity of A we
conclude
u0 − û0 + (ug − ûg ) = 0
and therefore uniqueness, u = u0 + ug = û0 + ûg .
52 3 Variational Methods
and therefore
1 0 1
u0 X ≤ A
f X + cA
2 ug X .
c1
Now the assertion follows from the triangle inequality and using assumption
(3.21).
is satisfied for all (v, q) ∈ X × Π. Note that for any solution (u, p) ∈ X × Π of
the extended variational problem (3.22) we conclude that u ∈ Vg is a solution
of Au = f . The second equation in (3.22) describes just the constraint u ∈ Vg
while the first equation in (3.22) coincides with the variational formulation
to find u0 ∈ V0 when choosing as test function v ∈ V0 . It remains to ensure
the existence of the Lagrange multiplier p ∈ Π such that the first equation in
(3.22) is satisfied for all v ∈ X, see Theorem 3.11.
3.5 Mixed Formulations 53
Proof. Let (u, p) be a solution of the variational problem (3.22). From the
first equation in (3.22) we then obtain
1
L(v, p) − L(u, p) = Av, v − f, v + Bv, p − g, p
2
1
− Au, u + f, u − Bu, p + g, p
2
1
= A(u − v), u − v + Au, v − u + B(v − u), p − f, v − u
2
1
= A(u − v), u − v ≥ 0,
2
and therefore
L(u, p) ≤ L(v, p) for all v ∈ X.
Using the second equation of (3.22) this gives
1
L(u, p) − L(u, q) = Au, u − f, u + Bu, p − g, p
2
1
− Au, u + f, u − Bu, q + g, q
2
= Bu, p − q − g, p − q = 0
and therefore
L(u, q) ≤ L(u, p) for all q ∈ Π.
For a fixed p ∈ Π we consider u ∈ X as the solution of the minimization
problem
L(u, p) ≤ L(v, p) for all v ∈ X.
Then we have for any arbitrary w ∈ X
d
L(u + tw, p)|t=0 = 0. (3.24)
dt
54 3 Variational Methods
From
1 1
L(u + tw, p) = Au, u − f, u + Bu, p − g, p + t2 Aw, w
2 2
+ t [Au, w + Bw, p − f, w] ,
0 ≤ L(u, p) − L(u, p + q)
1
= Au, u − f, u + Bu, p − g, p
2
1
− Au, u + f, u − Bu, p + q + g, p + q
2
= −Bu, q + g, q .
and therefore,
Bu, q = g, q for all q ∈ Π,
which is the second equation of (3.22).
Any solution (u, p) ∈ X × Π of the variational problem (3.22) is hence
a saddle point of the Lagrange functional L(·, ·). This is why the variational
problem (3.22) is often called a saddle point problem. The unique solvability
of (3.22) now follows from the following result.
Bv, q
cS qΠ ≤ sup for all q ∈ Π (3.25)
0=v∈X vX
is satisfied.
3.5 Mixed Formulations 55
and
1 cA
pΠ ≤ 1 + 2A f X + cB cA
2 gΠ . (3.27)
cS c1
Proof. Applying Theorem 3.8 we first find a unique u ∈ X satisfying
and
Bu, q = g, q for all q ∈ Π.
The estimate (3.26) is just the estimate of Corollary 3.9.
It remains to find p ∈ Π as the solution of the variational problem
and
Bv, p̂ = f − Au, v for all v ∈ X.
Then,
Bv, p − p̂ = 0 for all v ∈ X.
Using the stability condition (3.25) we obtain
Bv, p − p̂
0 ≤ cS p − p̂Π ≤ sup = 0
0=v∈X vX
and therefore p = p̂ in Π.
Using again (3.25) for the unique solution p ∈ Π this gives
Bv, p f − Au, v
cS pΠ ≤ sup = sup ≤ f X + cA
2 uX
0=v∈X vX 0=v∈X vX
and 2
1 [cB
2 ] cB
2
uX ≤ 1+ f X + gΠ . (3.31)
cA
1
A S
c1 c1 c1 cS1
A
and therefore
1 cB
2
uX ≤ f X + pΠ .
cA
1 cA
1
(I − K)u = 0
58 3 Variational Methods
(I − K)u = g
uX ≤ c gX .
uX ≤ c f X .
uX ≤ c D−1 f X ≤
c f X .
4
Variational Formulations
of Boundary Value Problems
In this chapter we describe and analyze variational methods for second order
elliptic boundary value problems as given in Chapter 1. To establish the unique
solvability of the associated variational formulations we will use the methods
which were given in the previous Chapter 3. The weak formulation of boundary
value problem is the basis to introduce finite element methods. Moreover,
from these results we can also derive mapping properties of boundary integral
operators (cf. Chapter 6) as used in boundary element methods.
d
∂ ∂
(Lu)(x) = − aji (x) u(x) for x ∈ Ω ⊂ Rd , (4.1)
i,j=1
∂xj ∂xi
d
∂
γ1int u(x) = lim nj (x)aji (
x) u(
x) for x ∈ Γ = ∂Ω. (4.2)
x→x∈Γ
Ω
i,j=1
∂
xi
|a(u, v)| ≤ cA
2 |u|H 1 (Ω) |v|H 1 (Ω) for all u, v ∈ H 1 (Ω) (4.6)
2 := d aL∞ (Ω) .
with cA
Proof. Using (4.5) we first have
d ∂ ∂
|a(u, v)| = aji (x) u(x) v(x)dx
i,j=1 ∂xi ∂xj
Ω
d
∂ d
∂
≤ aL∞ (Ω)
∂xi u(x) ∂xj v(x) dx.
Ω i=1 j=1
|a(u, v)| ≤ cA
2 uH 1 (Ω) vH 1 (Ω) for all u, v ∈ H 1 (Ω). (4.7)
We start to consider the Dirichlet boundary value problem (1.10) and (1.11),
Proof. For any given Dirichlet datum g ∈ H 1/2 (Γ ) we find, by applying the
inverse trace theorem (Theorem 2.22), a bounded extension ug ∈ H 1 (Ω) sat-
isfying γ0int ug = g and
Recall that
⎧⎡ ⎤2 ⎫1/2
⎪
⎨ ⎪
⎬
vW21 (Ω),Γ := ⎣ γ0 v(x) dsx ⎦ + ∇vL2 (Ω)
int 2
⎪
⎩ ⎪
⎭
Γ
defines an equivalent norm in H 1 (Ω) (cf. Example 2.7). For v ∈ V0 = H01 (Ω)
we then find from Lemma 4.2
w, zΓ
wH −1/2 (Γ ) = sup for all w ∈ H −1/2 (Γ ). (4.15)
0=z∈H 1/2 (Γ ) zH 1/2 (Γ )
4.1 Potential Equation 63
Lemma 4.4. Let u ∈ H 1 (Ω) be the unique solution of the Dirichlet boundary
value problem (4.10) when assuming g ∈ H 1/2 (Γ ) and f ∈ H −1 (Ω). For the
int −1/2
associated conormal derivative γ1 u ∈ H (Γ ) we then have
! "
γ1int uH −1/2 (Γ ) ≤ cIT f H
-−1 (Ω) + c2 |u|H 1 (Ω) .
A
(4.16)
Proof. Using the stability condition (4.15) and the variational formulation
(4.14) we find from the boundedness of the bilinear form a(·, ·) and by applying
the inverse trace theorem
|γ1int u, zΓ |
γ1int uH −1/2 (Γ ) = sup
0=z∈H 1/2 (Γ ) zH 1/2 (Γ )
Corollary 4.5. Let u ∈ H 1 (Ω) be the weak solution of the Dirichlet boundary
value problem
The assertion now follows from the semi–ellipticity (4.8) of the bilinear form
a(·, ·).
When Ω is a Lipschitz domain we can formulate stronger assumptions on
the given data f and g to establish higher regularity results for the solution
u of the Dirichlet boundary value problem and for the associated conormal
derivative γ1int u.
64 4 Variational Formulations of Boundary Value Problems
is satisfied for all (v, µ) ∈ H 1 (Ω) × H −1/2 (Γ ). Here we have used the bilinear
form
b(v, µ) := γ0int v, µΓ for (v, µ) ∈ H 1 (Ω) × H −1/2 (Γ ).
To investigate the unique solvability of the saddle point problem (4.18) we
will apply Theorem 3.11. Obviously,
! "
ker B := v ∈ H 1 (Ω) : γ0int v, µΓ = 0 for all µ ∈ H −1/2 (Γ ) = H01 (Ω).
Hence, due to (4.13), we have the ker B–ellipticity of the bilinear form a(·, ·).
It remains to establish the stability condition
γ0int v, µΓ
cS µH −1/2 (Γ ) ≤ sup for all µ ∈ H −1/2 (Γ ) . (4.19)
0=v∈H 1 (Ω) vH 1 (Ω)
4.1 Potential Equation 65
Lemma 4.7. The stability condition (4.19) is satisfied for all µ ∈ H −1/2 (Γ ).
On the other hand, with the Dirichlet boundary condition γ0int u = g we also
have
γ0int u(x)dsx = g(x)dsx . (4.21)
Γ Γ
Hence we can reformulate the saddle point problem (4.18) to find (u, λ) ∈
H 1 (Ω) × H −1/2 (Γ ) such that
66 4 Variational Formulations of Boundary Value Problems
γ0int u(x)dsx γ0int v(x)dsx + a(u, v) − b(v, λ) (4.22)
Γ Γ
= f, vΩ + g(x)dsx γ0int v(x)dsx
Γ Γ
b(u, µ) + λ(x)dsx µ(x)dsx = g, µΓ − f (x)dx µ(x)dsx (4.23)
Γ Γ Ω Γ
1 −1/2
is satisfied for all (v, µ) ∈ H (Ω) × H (Γ ).
The modified saddle point problem (4.22) and (4.23) is uniquely solv-
able, and the solution is also the unique solution of the original saddle point
problem (4.18), i.e. the saddle point formulations (4.22)–(4.23) and (4.18) are
equivalent.
Theorem 4.8. The modified saddle point problem (4.22) and (4.23) has a
unique solution (u, λ) ∈ H 1 (Ω) × H −1/2 (Γ ), which is also the unique solution
of the saddle point formulation (4.18).
Proof. The extended bilinear form
a(u, v) := γ0int u(x)dsx γ0int v(x)dsx + a(u, v)
Γ Γ
1
is bounded for all u, v ∈ H (Ω). Using Lemma 4.2 and Example 2.7 we find
⎡ ⎤2
a(v, v) = ⎣ γ0int v(x)dsx ⎦ + a(v, v) ≥ min{1, λ0 }v2 1
≥ cA 2
1 vH 1 (Ω) W2 (Ω),Γ
Γ
for all v ∈ H 1 (Ω) and therefore the H 1 (Ω)–ellipticity of the extended bilinear
form a(·, ·). Applying Theorem 3.11 we obtain as in Theorem 3.13 the unique
solvability of the saddle point problem (4.22) and (4.23). In particular for
(v, µ) ≡ (1, 1) we have
|Γ | γ0int u(x)dsx − λ(x)dsx = f (x)dx + |Γ | g(x)dsx ,
Γ Γ Ω Γ
γ0int u(x)dsx + |Γ | λ(x)dsx = g(x)dsx − |Γ | f (x)dx.
Γ Γ Γ Ω
Multiplying the first equation with |Γ | > 0 and adding the result to the second
equation this gives
2 int 2
(1 + |Γ | ) γ0 u(x)dsx = (1 + |Γ | ) g(x)dsx
Γ Γ
and therefore (4.21). Then we immediately get also (4.20), i.e. (u, λ) is also a
solution of the saddle point problem (4.18).
4.1 Potential Equation 67
Moreover, the solution of the Neumann boundary value problem (4.24) is only
unique up to an additive constant. To fix this constant, we formulate a suitable
scaling condition. For this we define
⎧ ⎫
⎨ ⎬
H∗1 (Ω) := v ∈ H 1 (Ω) : v(x)dx = 0 .
⎩ ⎭
Ω
Using Green’s first formula (4.3) we obtain the variational formulation of the
Neumann boundary value problem (4.24) to find u ∈ H∗1 (Ω) such that
defines an equivalent norm in H 1 (Ω) (cf. Example 2.7). Using Lemma 4.2 we
then have
for all v ∈ H∗1 (Ω) and therefore the H∗1 (Ω)–ellipticity of the bilinear form
a(·, ·) follows. The unique solvability of the variational problem (4.26) we
68 4 Variational Formulations of Boundary Value Problems
now conclude from Theorem 3.4 (Lax–Milgram lemma). Using the H∗1 (Ω)–
ellipticity of the bilinear form a(·, ·) we further have
2 int
1 uH 1 (Ω) ≤ a(u, u) = f, uΩ + g, γ0 uΓ
cA
≤ f H int
-−1 (Ω) uH 1 (Ω) + gH −1/2 (Γ ) γ0 uH 1/2 (Γ ) .
is satisfied for all v ∈ H 1 (Ω). To establish the unique solvability of the saddle
point problem (4.28) we have to investigate the assumptions of Theorem 3.11.
The bilinear form
b(v, µ) := µ v(x)dx for all v ∈ H 1 (Ω), µ ∈ R
Ω
is bounded, and we have ker B = H∗1 (Ω). Hence we obtain the ker B–ellipticity
of the bilinear form a(·, ·) from the ellipticity estimate (4.27). It remains to
prove the stability condition
b(v, µ)
cS |µ| ≤ sup for all µ ∈ R. (4.29)
0=v∈H 1 (Ω) vH 1 (Ω)
∗ 1
For an arbitrary given µ ∈ R we define . v := µ ∈ H (Ω) to obtain the
stability estimate (4.29) with cS = 1/ |Ω|. By applying Theorem 3.11 we
now conclude the unique solvability of the saddle point problem (4.28).
Choosing in (4.28) the test function v ≡ 1 we obtain for the Lagrange
parameter λ from the solvability condition (4.25)
λ = 0.
Instead of (4.28) we may now consider an equivalent saddle point formulation
to find (u, λ) ∈ H 1 (Ω) × R such that
a(u, v) + λ v(x)dx = f, vΩ + g, γ0int vΓ
Ω (4.30)
u(x)dx − λ =0
Ω
4.1 Potential Equation 69
is satisfied for all v ∈ H 1 (Ω). Using the second equation we can eliminate the
scalar Lagrange multiplier λ ∈ R to obtain a modified variational problem to
find u ∈ H 1 (Ω) such that
a(u, v) + u(x)dx v(x)dx = f, vΩ + g, γ0int vΓ (4.31)
Ω Ω
Theorem 4.10. For any f ∈ H −1 (Ω) and for any g ∈ H −1/2 (Γ ) there is a
1
unique solution u ∈ H (Ω) of the modified variational problem (4.31).
If f ∈ H −1 (Ω) and g ∈ H −1/2 (Γ ) satisfy the solvability condition (4.25),
then we have u ∈ H∗1 (Ω), i.e. the modified variational problem (4.31) and the
saddle point formulation (4.28) are equivalent.
and therefore u ∈ H∗1 (Ω). The solution of the modified variational problem
(4.31) is therefore also a solution of the saddle point formulation (4.28), i.e.
both formulations are equivalent.
Since the solution of the Neumann boundary value problem (4.24) is not
unique, we can add an arbitrary constant α ∈ R to the solution u ∈ H∗1 (Ω)
to obtain the general solution u := u + α ∈ H 1 (Ω).
70 4 Variational Formulations of Boundary Value Problems
Proof. For gD ∈ H 1/2 (ΓD ) we first find a bounded extension gD ∈ H 1/2 (Γ )
satisfying
gD H 1/2 (Γ ) ≤ c gD H 1/2 (ΓD ) .
Applying the inverse trace theorem (Theorem 2.22) there exists a second ex-
tension ugD ∈ H 1 (Ω) with γ0int ugD = gD and satisfying
We finally consider the boundary value problem (1.10) and (1.13) with bound-
ary conditions of Robin type,
d
∂
Li u(x) = − σij (u, x) for x ∈ Ω ⊂ Rd , i = 1, . . . , d.
j=1
∂x j
where we assume E > 0 and ν ∈ (0, 1/2). Using the associated conormal
derivative (1.23) and the bilinear form (1.26),
d
a(u, v) = 2 µ eij (u, x)eij (v, x)dx + λ div u(x) div v(x)dx
Ω i,j=1 Ω
d
= σij (u, x)eij (v, x)dx,
Ω i,j=1
or in short,
E
σ = C e.
(1 + ν)(1 − 2ν)
Due to the symmetries σij (u, x) = σji (u, x) and eij (v, x) = eji (v, x) we have
E
a(u, v) = (DCe(u, x), e(v, x)) dx
(1 + ν)(1 − 2ν)
Ω
with the diagonal matrix D = diag(1, 1, 1, 2, 2, 2). The eigenvalues of the ma-
trix DC ∈ R6×6 are
Using
d d 2
1 ∂
d d
∂
e(v, x)22 = 2
[eij (u, x)] = uj (x) + ui (x)
i=1 j=i
4 i=1 j=i ∂xi ∂xj
2 2 + d 2
1
d
∂ ∂ ∂
≤ ui (x) + uj (x) = ui (x)
2 i,j=1 ∂xj ∂xi i,j=1
∂xj
we obtain
74 4 Variational Formulations of Boundary Value Problems
d 2
∂
e(u, x)22 dx ≤ ui (x) dx = |u|2[H 1 (Ω)]d
i,j=1
∂x j
Ω Ω
and therefore the estimate (4.36). In the case d = 2 the assertion follows in
the same way.
The proof of the [H01 (Ω)]d –ellipticity of the bilinear form a(·, ·) requires
several steps.
d 2 d
1 ∂ 1 ∂ ∂
= ϕi (x) dx + ϕi (x) ϕj (x)dx.
2 i,j=1 ∂xj 2 i,j=1 ∂xj ∂xi
Ω Ω
and therefore
d
d
∂ ∂ ∂ ∂
ϕi (x) ϕj (x)dx = ϕi (x) ϕj (x)dx
i,j=1 Ω
∂xj ∂xi i,j=1
∂xi ∂xj
Ω
#d
$2
∂
= ϕi (x) dx ≥ 0.
i=1
∂xi
Ω
Hence we have
d 2
1
d
2 ∂ 1
[eij (ϕ, x)] dx ≥ ϕi (x) dx = |ϕ|2[H 1 (Ω)]d .
i,j=1
2 i,j=1
∂x j 2
Ω Ω
Considering the closure of C0∞ (Ω) with respect to the norm · H 1 (Ω) we
conclude the assertion for v ∈ [H01 (Ω)]d .
Using suitable equivalent norms in [H 1 (Ω)]d we now conclude the [H01 (Ω)]d –
ellipticity of the bilinear form a(·, ·).
an equivalent norm in [H 1 (Ω)]d (cf. Theorem 2.6). The assertion then follows
from Lemma 4.14 and by using Korn’s first inequality (4.38).
The ellipticity estimate (4.39) remains valid for vector functions v, where
only some components vi (x) are zero for x ∈ ΓD,i ⊂ Γ . Let
! "
[H01 (Ω, ΓD )]d = v ∈ [H 1 (Ω)]d : γ0int vi (x) = 0 for x ∈ ΓD,i , i = 1, . . . , d .
Then we have
As for the scalar Laplace operator we can extend the bilinear form a(·, ·)
of the system of linear elasticity by some L2 norm to obtain an equivalent
norm in [H 1 (Ω)]d . This is a direct consequence of Korn’s second inequality,
see [53].
76 4 Variational Formulations of Boundary Value Problems
For the solution u ∈ [H 1 (Ω)]d of the Dirichlet boundary value problem we now
compute the associated boundary stress γ1int u ∈ [H −1/2 (Γ )]d as the solution
of the variational problem
γ1int u, wΓ = a(u, Ew) − f , EwΩ
for all w ∈ [H 1/2 (Γ )]d . Here, E : H 1/2 (Γ ) → H 1 (Ω) is the extension operator
which is applied to the components wi ∈ H 1/2 (Ω). Note that
E
γ1int u[H −1/2 (Γ )]d ≤ cIT f [H -−1 (Ω)]d + |u| 1 .
1 − 2ν [H (Ω)]
d
4.2 Linear Elasticity 77
Lemma 4.19. Let u ∈ [H 1 (Ω)]d be the weak solution of the Dirichlet bound-
ary value problem
Then we have
a(u, u) ≥ c γ1int u2[H −1/2 (Γ )]d . (4.43)
and
e(Ew, x)22 dx ≤ |Ew|2[H 1 (Ω)]d .
Ω
Moreover,
d
1+ν
e(u, x)22 dx ≤ [eij (u, x)]2 dx ≤ a(u, u).
E
Ω Ω i,j=1
γ1int u, wΓ
γ1int u[H −1/2 (Γ )]d = sup
0=w∈[H 1/2 (Γ )]d w[H 1/2 (Γ )]d
a(u, Ew) .
= sup ≤ c a(u, u).
0=w∈[H 1/2 (Γ )]d w[H 1/2 (Γ )]d
Note that the constant c in the estimate (4.43) tends to zero when ν → 12 .
where v k are the rigid body motions (cf. (1.29)). On the other hand, the
solution of the Neumann boundary value problem is only unique up to the
rigid body motions. To fix the rigid body motions, we formulate appropriate
scaling conditions. For this we define
⎧ ⎫
⎨ ⎬
[H∗1 (Ω)]d = v ∈ [H 1 (Ω)]d : v k (x) v(x)dx = 0 for all v k ∈ R .
⎩ ⎭
Ω
is satisfied for all v ∈ [H∗1 (Ω)]d . Using Corollary 4.18 we can establish the
[H∗1 (Ω)]d –ellipticity of the bilinear form a(·, ·) and therefore we can conclude
the unique solvability of the variational problem (4.44) in [H∗1 (Ω)]d .
By introducing Lagrange multipliers we can formulate the scaling condi-
tions conditions of [H∗1 (Ω)]d as side conditions in a saddle point problem.
Then we have to find u ∈ [H 1 (Ω)]d and λ ∈ Rdim(R) such that
R
dim
a(u, v) + λk v k (x) v(x)dx = f , vΩ + g, γ0int vΓ
k=1 Ω (4.45)
v (x) u(x)dx =0
Ω
R
dim
λk v k (x) v (x)dx = 0 for = 1, . . . , dim(R).
k=1 Ω
Since the rigid body motions are linear independent, we obtain λ = 0. In-
serting this result into the second equation in (4.45), and eliminating the
Lagrange multiplier λ we finally obtain a modified variational problem to find
u ∈ [H 1 (Ω)]d such that
R
dim
a(u, v) + v k (x) u(x)dx v k (x) v(x)dx = f , vΩ + g, γ0int vΓ
k=1 Ω Ω
(4.46)
is satisfied for all v ∈ [H 1 (Ω)]d .
4.3 Stokes Problem 79
R
dim
:= u +
u αk v k ∈ [H 1 (Ω)]d
k=1
d
a(u, v) = f , vΩ + gN,i , γ0int vi ΓN,i (4.47)
i=1
Next we consider the Dirichlet boundary value problem for the Stokes system
(1.38),
for all v ∈ [H01 (Ω)]d . Using the equivalent norm (4.40) we then find the
[H01 (Ω)]d –ellipticity of the bilinear form a(·, ·),
a(v, v) ≥ c v2[H 1 (Ω)]d for all v ∈ [H01 (Ω)]d .
Due to ker B ⊂ [H01 (Ω)]d we also have the ker B–ellipticity of the bilinear
form a(·, ·). It remains to prove the stability condition (3.25),
q(x) div v(x)dx
Ω
cS qL2 (Ω) ≤ sup for all q ∈ L2,0 (Ω). (4.51)
0=v∈[H01 (Ω)]d v[H 1 (Ω)]d
This is a direct consequence of Theorem 2.17:
Lemma 4.20. Let Ω ⊂ Rd be a bounded and connected Lipschitz domain.
Then there holds the stability condition (4.51) .
Proof. For q ∈ L2,0 (Ω) we have ∇q ∈ [H −1 (Ω)]d satisfying, by using Theorem
2.17,
qL2 (Ω) ≤ c ∇q[H −1 (Ω)]d .
Recalling the norm definition in [H −1 (Ω)]d by duality, this gives
1 w, ∇qΩ
qL2 (Ω) ≤ ∇q[H −1 (Ω)]d = sup
c 1
0=w∈[H0 (Ω)] d w[H 1 (Ω)]d
− q(x) div w(x)dx
Ω
= sup .
0=w∈[H01 (Ω)]d w[H 1 (Ω)]d
Hence, choosing v := −w we finally obtain the stability condition (4.51).
82 4 Variational Formulations of Boundary Value Problems
Therefore, all assumptions of Theorem 3.11 are satisfied, and hence, the saddle
point problem (4.50) is unique solvable.
The scaling condition to fix the pressure p ∈ L2,0 (Ω) can now be reformu-
lated as for the Neumann boundary value problem for the potential equation.
By introducing a scalar Lagrange multiplier λ ∈ R we may consider the fol-
lowing extended saddle point problem to find u ∈ [H 1 (Ω)]d with u(x) = g(x)
for x ∈ Γ as well as p ∈ L2 (Ω) and λ ∈ R such that
a(u, v) − p(x) div v(x)dx = f , vΩ ,
Ω
q(x) div u(x) dx + λ q(x)dx = 0, (4.52)
Ω Ω
p(x)dx =0
Ω
is satisfied for all v ∈ [H01 (Ω)]d and q ∈ L2 (Ω). Choosing as test function
q ≡ 1 this gives
λ |Ω| = − div u(x) dx = − n(x) g(x)dsx = 0
Ω Γ
and using the solvability condition (4.48) we get λ = 0. Hence we can write
the third equation in (4.52) as
p(x)dx − λ = 0.
Ω
induces f ∈ [H −1 (Ω)]d .
4.3 Stokes Problem 83
By
for all u, v ∈ [H01 (Ω)]d and p, q ∈ L2 (Ω) we can define bounded operators
and inserting this into the second equation we obtain the Schur complement
system 0 1
BA−1 B + D p = −Bug − A−1 f. (4.54)
and therefore
cB
2
up [H 1 (Ω)]d ≤ pL2 (Ω) .
cA
1
84 4 Variational Formulations of Boundary Value Problems
≤ c pL2 (Ω) qL2 (Ω)
For p0 ∈ L2,0 (Ω) we can use the stability condition (4.51), the definition of
up0 = A−1 B p0 ∈ [H01 (Ω)]d and the boundedness of A to obtain
b(v, p0 )
cS p0 L2 (Ω) ≤ sup
0=v∈[H01 (Ω)]d v[H 1 (Ω)]d
a(up0 , v)
= sup ≤ cA
2 up0 [H 1 (Ω)]d
0=v∈[H01 (Ω)]d v[H 1 (Ω)]d
and therefore
2 2
cA
2 1 cA
2
p0 2L2 (Ω) ≤ up0 2[H 1 (Ω)]d ≤ a(up0 , up0 ) = c b(up0 , p0 ).
cS cA
1 cS
1
BA−1 B p0 , p0 L2 (Ω) = Bup0 , p0 L2 (Ω) ≥ p0 2L2 (Ω) .
c
For v ∈ [H01 (Ω)]d we have
B p, vΩ = p(x)div v(x)dx = −v, ∇pΩ = −v, ∇p0 Ω
Ω
4.4 Helmholtz Equation 85
and hence
1
BA−1 B p, pL2 (Ω) = BA−1 B p0 , p0 L2 (Ω) ≥ p0 2L2 (Ω) .
c
From this we obtain
i.e., S is L2 (Ω)–elliptic.
Applying Theorem 3.4 (Lax–Milgram lemma) we finally obtain the unique
solvability of the operator equation (4.54).
The related variational formulation is to find u ∈ H 1 (Ω) with γ0int u(x) = g(x)
for x ∈ Γ such that
∇u(x)∇v(x)dx − k 2 u(x)v(x)dx = 0 (4.56)
Ω Ω
can be written as
a(u, v) = a0 (u, v) − c(u, v) (4.57)
where the symmetric and bounded bilinear form
a0 (u, v) = ∇u(x)∇v(x) dx for u, v ∈ H 1 (Ω)
Ω
86 4 Variational Formulations of Boundary Value Problems
4.5 Exercises
4.1 Derive the variational formulation of the following boundary value prob-
lem with nonlinear Robin boundary conditions
where
ε for x ∈ Ω0 ⊂ Ω,
α(x) =
1 for x ∈ Ω\Ω0 .
Prove the unique solvability of the related variational formulation. Discuss the
dependence of the constants on the parameter ε << 1. Can these constants be
improved when using other norms?
4.4 Formulate a sufficient condition on the wave number k such that the
variational formulation of the interior Neumann boundary value problem of
the Helmholtz equation,
d
∂ ∂
(Lu)(x) = − aji (x) u(x) .
i,j=1
∂xj ∂xi
d
∂
γ1int u(x) = nj (x)aji (x) u(x) for x ∈ Γ
i,j=1
∂xi
and Green’s second formula (1.8) reads for the solution u of the partial dif-
ferential equation (1.10) and for an arbitrary test function v
(Lv)(y)u(y)dy = γ1 u(y)γ0 v(y)dsy − γ1int v(y)γ0int u(y)dsy
int int
Ω Γ Γ
+ f (y)v(y)dy.
Ω
then the solution of the partial differential equation (1.10) is given by the
representation formula for x ∈ Ω
90 5 Fundamental Solutions
u(x) = U ∗ (x, y)γ1int u(y)dsy − int U ∗ (x, y)γ int u(y)ds
γ1,y 0 y (5.2)
Γ Γ
+ U ∗ (x, y)f (y)dy.
Ω
Hence we can describe any solution of the partial differential equation (1.10)
just by knowing the Cauchy data [γ0int u(x), γ1int u(x)] for x ∈ Γ .
Due to
u(x) = δ0 (y − x)u(y)dy for x ∈ Ω
Ω
Since the Laplace operator is invariant with respect to translations and rota-
tions, we can find the fundamental solution as U ∗ (x, y) = v(z) with z := y −x.
Hence we have to solve
Applying the Fourier transformation (2.14) this gives, when considering the
derivation rule (2.17),
5.1 Laplace Operator 91
1
|ξ|2 v%(ξ) =
(2π)d/2
and therefore
1 1
v%(ξ) = ∈ S (Rd ).
(2π)d/2 |ξ|2
For the Fourier transform v% of a tempered distribution v ∈ S (Rd ) we have
by definition
%
v , ϕL2 (Rd ) = v, ϕ
% L2 (Rd ) for all ϕ ∈ S(Rd ).
Using
ϕ(ξ) = (2π)−d/2 %
eiz,ξ
ϕ(z)dz
Rd
it follows that
1 1
%
v , ϕL2 (Rd ) = %
eiz,ξ
ϕ(z)dzdξ.
(2π)d |ξ|2
Rd Rd
does not exist we can not exchange the order of integration. However, using
for r ∈ (0, ∞), ϕ ∈ (0, 2π), θ ∈ (0, π) to obtain, by using Lemma 2.13,
⎡ ⎤
iz,ξ
iz,ξ
1 ⎢ e e ⎥
v(z) = v(|z|) = ⎣ dξ − ∆z dξ ⎦
(2π)3 |ξ|2 |ξ|4
|ξ|≤1 |ξ|>1
⎡ 2π π 1
1 ⎣
= ei|z|r cos θ sin θ dr dθ dϕ
(2π)3
0 0 0
⎤
2ππ ∞
ei|z|r cos θ sin θ
−∆z dr dθ dϕ⎦
r2
0 0 1
⎡ π 1 ⎤
π ∞ i|z|r cos θ
1 ⎣ e sin θ
= ei|z|r cos θ sin θ dr dθ − ∆z dr dθ ⎦ .
(2π)2 r2
0 0 0 1
π 1
i|z|r cos θ 1 2 i|z|r 3 2
e sin θdθ = ei|z|ru du = e − e−i|z|r = sin |z|r
i|z|r |z|r
0 −1
and therefore
⎡ ⎤
1 ∞
1 ⎣ sin |z|r sin |z|r ⎦
v(z) = dr − ∆z dr .
2π 2 |z|r |z|r3
0 1
1 |z|
sin |z|r 1 sin s Si(|z|)
I1 := dr = ds = .
|z|r |z| s |z|
0 0
With
sin ax 1 sin ax a cos ax
dx = − + dx
x3 2 x2 2 x2
1 sin ax a cos ax a2 sin ax
=− − − dx
2 x2 2 x 2 x
the computation of the second integral gives
5.1 Laplace Operator 93
∞ ∞ ∞
sin |z|r 1 sin |z|r 1 cos |z|r |z| sin |z|r
I2 := 3
dr = − 2
− − dr
|z|r 2 |z|r 2 r 1 2 r
1 1
⎡∞ ⎤
1
1 sin |z| 1 |z| ⎣ sin |z|r sin |z|r ⎦
= + cos |z| − dr − dr
2 |z| 2 2 r r
0 0
1 sin |z| 1 |z| 2 π 3
= + cos |z| − − Si(|z|) .
2 |z| 2 2 2
Hence the fundamental solution of the Laplace operator in three space dimen-
sions is
1 1
U ∗ (x, y) = for x, y ∈ R3 .
4π |x − y|
For the two–dimensional case d = 2 the inverse Fourier transform of the
fundamental solution has to be regularized in some appropriate way [154]. By
1 ϕ(x) − ϕ(0) ϕ(x)
P 2 , ϕL2 (Rd ) = dx + dx
|x| |x|2 |x|2
x∈R2 :|x|≤1 x∈R2 :|x|>1
1
we first define the tempered distribution P ∈ S (Rd ). Then,
|x|2
1 ϕ(ξ) − ϕ(0) ϕ(ξ)
2π v, ϕ
% L2 (R2 ) = P , ϕL2 (R2 ) = dξ + dξ
|ξ|2 |ξ|2 |ξ|2
ξ∈R2 :|ξ|≤1 ξ∈R2 :|ξ|>1
we then obtain
1 1
(2π)2 v, ϕ
% L2 (R2 ) = [eiz,ξ
− 1]ϕ(z)dzdξ
% + %
eiz,ξ
ϕ(z)dzdξ.
|ξ|2 |ξ|2
ξ∈R2 :|ξ|≤1 R2 ξ∈R2 :|ξ|>1 R2
94 5 Fundamental Solutions
Again we can not exchange the order of integration in the second term. How-
ever, as in three–dimensional case we can write
⎡ ⎤
⎢ eiz,ξ
− 1 eiz,ξ
⎥
(2π)2 v, ϕ
% L2 (R2 ) = ϕ(z)
% ⎣ 2
dξ + dξ ⎦ dz.
|ξ| |ξ|2
R2 ξ∈R2 :|ξ|≤1 ξ∈R2 :|ξ|>1
Since any constant satisfies the homogeneous Laplace equation we can ne-
glect constant terms in the definition of the fundamental solution. Hence the
fundamental solution of the Laplace operator in two space dimensions is
5.1 Laplace Operator 95
1
U ∗ (x, y) = − log |x − y| for x, y ∈ R2 .
2π
In what follows we will describe an alternative approach to compute the fun-
damental solution for the Laplace operator in two space dimensions. From
Lemma 2.13 we know that the solution v(z) depends only on the absolute
value := |z|. For z = 0 the partial differential equation (5.4) can be rewrit-
ten in polar coordinates as
2
∂ 1 ∂
+ v() = 0 for > 0.
∂2 ∂
v() = a log + b, a, b ∈ R.
U ∗ (x, y) = a log |x − y| .
For x ∈ Ω and for a sufficient small ε > 0 let Bε (x) ⊂ Ω be a ball with center
x and with radius ε. For y ∈ Ω\Bε (x) the fundamental solution U ∗ (x, y) is
a solution of the homogeneous Laplace equation −∆y U ∗ (x, y) = 0. Applying
Green’s second formula (1.8) with respect to the bounded domain Ω\Bε (x)
we obtain
∂ ∂
0 = U ∗ (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy + U ∗ (x, y)f (y)dy
∂ny ∂ny
Γ Γ Ω\B ε (x)
∂ ∂
+ U ∗ (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy .
∂ny ∂ny
∂Bε (x) ∂Bε (x)
1
Using ny = (x − y) for y ∈ ∂Bε (x) we have
ε
∂ ∗ (ny , y − x) a
U (x, y)u(y)dsy = a u(y)dsy = − u(y)dsy .
∂ny |x − y|2 ε
∂Bε (x) ∂Bε (x) ∂Bε (x)
where
a
(y − x)∇u(ξ)dy ≤ |a| 2π ε uC 1 (Ω) .
ε
∂Bε (x)
Taking the limit ε → 0 gives
∗ ∂ ∂
−a 2π u(x) = U (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy
∂ny ∂ny
Γ Γ
+ U ∗ (x, y)f (y)dy
Ω
Γ Ω
Using (1.32),
λ+µ
v k (z) := ∆[ψ(z)ek ] − grad div [ψ(z)ek ],
λ + 2µ
we have to find the Airy stress function ψ satisfying the Bi–Laplace equation
or
−µ∆ϕ(z) = δ0 (z), ∆ψ(z) = ϕ(z) for z ∈ Rd .
From the fundamental solution of the Laplace operator we find
⎧
⎪ 1 1
⎪
⎨ − µ 2π log |z|,
⎪ for d = 2,
ϕ(z) =
⎪
⎪ 1 1 1
⎪
⎩ for d = 3.
µ 4π |z|
For d = 2 we have to solve the remaining Poisson equation when using polar
coordinates,
2
∂ 1 ∂ 1 1
2
+ ψ() = − log for > 0,
∂ ∂ µ 2π
1 1 0 2 1
ψ() = − log − 2 + a log + b for > 0, a, b ∈ R.
µ 8π
In particular for a = b = 0 we have
1 1 0 2 1
ψ() = − log − 2 .
µ 8π
λ + µ ∂2 λ + µ ∂2
v11 (z) = ∆ψ(z) − ψ(z), v21 (z) = − ψ(z).
λ + 2µ ∂z12 λ + 2µ ∂z1 ∂z2
∂ zi
Using |z| = we obtain
∂zi |z|
∂ 1 1
ψ(z) = − [2zi log |z| + zi ] (i = 1, 2)
∂zi µ 8π
∂2 1 1 z12
ψ(z) = − 2 log |z| + 2 + 1 (i = 1, 2)
∂zi2 µ 8π |z|2
∂2 1 1 z1 z 2
ψ(z) = −
∂z1 ∂z2 µ 4π |z|2
and therefore
2
1 λ + 3µ 1 λ+µ z1 3
v11 (z) = − log |z| + − ,
4π µ(λ + 2µ) 4π µ(λ + 2µ) |z|2 2
1 λ + µ z1 z 2
v21 (z) = .
4π µ(λ + 2µ) |z|2
For k = 2 the computation is almost the same. Since the constants are so-
lutions of the homogeneous system we can neglect them when defining the
fundamental solution. From v k for k = 1, 2 we then find the Kelvin solution
tensor U ∗ (x, y) = (v 1 , v 2 ) with the components
∗ 1 λ+µ λ + 3µ (yk − xk )(y − x )
Uk (x, y) = − log |x − y| δk +
4π µ(λ + 2µ) λ+µ |x − y|2
λ + µ ∂2
v11 (z) = ∆ψ(z) − ψ(z),
λ + 2µ ∂z12
λ + µ ∂2
v21 (z) = − ψ(z),
λ + 2µ ∂z1 ∂z2
λ+µ ∂
v31 (z) = − ψ(z)
λ + 2µ ∂z1 ∂z3
and with the derivatives
∂ 1 1 zi ∂2 1 1 1 z2
ψ(z) = , 2 ψ(z) = − i3 ,
∂zi µ 8π |z| ∂zi µ 8π |z| |z|
∂2 1 1 zi zj
ψ(z) = − for i = j
∂zi ∂zj µ 8π |z|3
for k, = 1, . . . , 3. Inserting the Lamé constants (1.24) this gives the funda-
mental solution of linear elastostatics in three space dimensions
∗ 1 1 1+ν δk (yk − xk )(y − x )
Uk (x, y) = (3 − 4ν) + .
8π E 1 − ν |x − y| |x − y|3
Hence we have the fundamental solution of linear elastostatics
∗ 1 1 1+ν (xk − yk )(x − y )
Uk (x, y) = (3 − 4ν)E(x, y)δk +
4(d − 1)π E 1 − ν |x − y|d
(5.9)
for k, = 1, . . . , d with
⎧
⎨ − log |x − y|
⎪ for d = 2,
E(x, y) = 1
⎪
⎩ for d = 3.
|x − y|
Inserting the solution vectors v(y) = U ∗k (x, y) into the second Betti formula
(5.8) this gives the representation formula
uk (x) = U ∗k (x, y) γ1int u(y)dsy − u(y) γ1,yint U ∗ (x, y)ds
k y
Γ Γ
+ f (y) U ∗k (x, y)dy (5.10)
Ω
T ∗k (x, y) := γ1,y
int U ∗ (x, y)
k
∂
= λ divy U ∗k (x, y) n(y) + 2µ U ∗ (x, y) + µ n(y) × curly U ∗k (x, y).
∂ny k
Using
1 1 1+ν yk − xk
div U ∗k (x, y) = 2(2ν − 1)
4(d − 1)π E 1 − ν |x − y|d
we then obtain
1 ν yk − xk E ∂
T ∗k (x, y) = − n(y) + U ∗ (x, y)
2(d − 1)π 1 − ν |x − y|d 1 + ν ∂ny k
E
+ n(y) × curly U ∗k (x, y). (5.11)
2(1 + ν)
Obviously, both the fundamental solutions U ∗k (x, y) and the corresponding
boundary stress functions T ∗k (x, y) exist also for incompressible materials with
ν = 1/2.
5.3 Stokes Problem 101
For the solution u and for an arbitrary vector field v we obtain from Green’s
first formula (1.41) by using the symmetry a(u, v) = a(v, u) Green’s second
formula
d
∂
−µ∆vi (y) + q(y) ui (y)dy + p(y)div v(y)dy (5.12)
∂yi
Ω i=1 Ω
d
d
= ti (u, p)vi (y)dsy − ti (v, q)ui (y)dsy + f (y) v(y)dy
Γ i=1 Γ i=1 Ω
1
d
µ |ξ|2 v%jk (ξ) + i ξj q%k (ξ) = δjk (j = 1, . . . , d), i ξj v%jk (ξ) = 0.
(2π)d/2 j=1
implies
1 1 1 0 2 1 C0 2
eiz,ξ
dξ = |z| log |z| − |z|2 + |z| + C1 + C2 log |z|
(2π)2 |ξ|4 8π 8π
R2
For d = 2 and k = 2 the computations are almost the same. Neglecting the
constants we finally have the fundamental solution for the Stokes system in
two space dimensions,
∗ 1 1 (yk − xk )(y − x )
Uk (x, y) = − log |x − y| δk + (5.13)
4π µ |x − y|2
1 yk − xk
Q∗k (x, y) = (5.14)
2π |x − y|2
and k, = 1, 2.
For d = 3 we obtain in the same way the fundamental solution for the
Stokes system as
∗ 1 1 δk (yk − xk )(y − x )
Uk (x, y) = +
8π µ |x − y| |x − y|3
1 yk − xk
Q∗k (x, y) =
4π |x − y|3
and k, = 1, . . . , 3.
Comparing the above results with the fundamental solution of the system
of linear elasticity we obtain the equality for
1 1+ν 1
= , (3 − 4ν) = 1
E 1−ν µ
and therefore for
1
ν =
, E = 3µ .
2
The fundamental solution of the linear elasticity system with incompressible
material therefore coincides with the fundamental solution of the Stokes sys-
tem.
Inserting the fundamental solutions v(y) = U ∗k (x, y) and q(y) = Q∗k (x, y)
into the second Greens formula (5.12) this gives the representation formulae
uk (x) = U k (x, y) t(u(y), p(y))dsy − u(y) t(U ∗k (x, y), Q∗k (x, y))dsy
∗
Γ Γ
+ f (y) U ∗k (x, y)dy (5.15)
Ω
Hence the boundary stress (1.43) of the fundamental solution of the Stokes
system also coincides with the boundary stress (1.27) of the fundamental
solution of the linear elasticity system when choosing ν = 12 and E = 3µ.
It remains to find some appropriate representation formulae for the pres-
sure p. Let us first consider the case d = 2 and the second Green formula
(5.12) where we have to find solutions v 3 (z) and q 3 (z) with z := y − x such
that
−µ∆v 3 (z) + ∇q 3 (z) = 0, div v 3 (z) = δ0 (z) for z ∈ R2 .
By applying the Fourier transformation we obtain the linear system
As before we obtain
1 ∂
vi3 (z) = log |z| (i = 1, 2), q 3 (z) = µδ0 (z).
2π ∂zi
Using z := y − x we conclude for x ∈ Ω a representation formula for the
pressure
2
2
p(x) = ti (u, p)vi3 (x, y)dsy − ti (v 3 (x, y), q 3 (x, y))ui (y)dsy
Γ i=1 Γ i=1
2
+ v 3i (x, y)fi (y)dy
Ω i=1
2 2
∂ 3 1 ∂2
div v 3 (x, y) = vi (x, y) = log |x − y| = 0,
i=1
∂yi 2π i=1 ∂yi2
5.4 Helmholtz Equation 105
we obtain for Γ y = x ∈ Ω
2
ti (v 3 (x, y), q 3 (x, y)) = 2µ eij (v 3 (x, y), y)nj (y).
j=1
Moreover,
3 1 ∂ 3 ∂ 3
eij (v (x, y), y) = v (x, y) + v (x, y)
2 ∂yi j ∂yj i
1 ∂ ∂ ∂ ∂
= log |x − y| + log |x − y|
4π ∂yi ∂yj ∂yj ∂yi
1 ∂ ∂
= log |x − y|
2π ∂yi ∂yj
1 ∂ ∂ ∂ ∗
=− log |x − y| = − Q (x, y).
2π ∂xj ∂yi ∂xj i
1 cos k|x − y|
Uk∗ (x, y) = for x, y ∈ R3 .
4π |x − y|
1 eik|x−y|
Uk∗ (x, y) = for x, y ∈ R3 . (5.18)
4π |x − y|
For d = 2, and by using polar coordinates the Helmholtz equation (5.17) reads
∂2 1 ∂
− v() − v() − k 2 v() = 0 for > 0,
∂2 ∂
or
∂2 ∂
2 v() + v() + k 2 2 v() = 0 for > 0.
∂2 ∂
With the substitution
s 1 ∂
s = k, v() = v( ) = V (s), V (s) = v()
k k ∂
we then obtain a Bessel differential equation of order zero,
and thus
1
v1 = 0, vk = − vk−2 for k ≥ 2.
k2
Hence we obtain
1
v2 −1 = 0, v2 = − v2( −1) for = 1, 2, . . .
42
and therefore
(−1)
v2 = v0 for = 1, 2, . . . .
4 (!)2
In particular for v0 = 1 we have
∞
(−1) 2
V1 (s) = 1 + s =: J0 (s)
4 (!)2
=1
which is the first kind Bessel function of order zero. Note that
lim J0 (s) = 1.
s→0
By using
1
V2 (s) = J0 (s) ln s + J0 (s) + W (s),
s
2 1
V2 (s) = J0 (s) ln s + J0 (s) − 2 J0 (s) + W (s)
s s
we obtain
With
∞
∞
W (s) = wk sk , J0 (s) = vk ksk−1
k=0 k=1
we have to solve
∞
∞
0 2 1
0= k wk + wk−2 sk + w1 s + 2 vk ksk
k=2 k=1
∞
0 2 1
= k wk + wk−2 + 2kvk sk + [w1 + 2v1 ] s for s > 0.
k=2
Hence we find
w1 = −2v1 = 0,
and
k 2 wk + wk−2 + 2kvk = 0 for k ≥ 2,
i.e.
1
wk = − [wk−2 + 2kvk ] for k ≥ 2.
k2
By using v2 −1 = 0 for ∈ N we then obtain w2 −1 = 0 for ∈ N and
1 0 1 1 1 (−1)
w2 = − w2( −1) + 4v2 = − w2( −1) − .
42 42 4 (!)2
(−1) +1 1
w2 = for ∈ N.
4 (!)2 j=1 j
Hence we have
Instead of V2 (s) we will use a linear combination of V1 (s) and V2 (s) to define
a second solution of the fundamental system, in particular we introduce the
second kind Bessel function of order zero,
where
5.5 Exercises 109
⎡ ⎤
n
1
γ = lim ⎣ − ln n⎦ ≈ 0.57721566490 . . .
n→∞
j=1
j
5.5 Exercises
5.1 Consider the recursion
1 1 (−1)
w0 = 0, w2 = − w2( −1) − for ∈ N.
42 4 (!)2
(−1) +1 1
w2 = for ∈ N.
4 (!)2 j=1 j
1
u(x) = G(x, y)f (y)dy for x ∈ (0, 1)
0
N 0 ψΩ
0 f, ψΩ := f, N for all ψ ∈ S(Rd ).
% is
where the Fourier transform ϕ
= (2π)− 2 e−ix,ξ
ϕ(x)dx.
d
%
ϕ(ξ)
Rd
Moreover,
0 ϕ)(x) =
u(x) := (N U ∗ (x, y)ϕ(y)dy for x ∈ Rd .
Ω
Let Ω ⊂ BR (0), and let µ ∈ C0∞ ([0, ∞)) be a non–negative, monotone decreas-
ing cut off function with compact support, and let µ(r) = 1 for r ∈ [0, 2R].
Define
uµ (x) := µ(|x − y|)U ∗ (x, y)ϕ(y)dy for x ∈ Rd .
Ω
and therefore
uH 1 (Ω) = uµ H 1 (Ω) ≤ uµ H 1 (Rd )
with
uµ 2H 1 (Rd ) = (1 + |ξ|2 ) |%
uµ (ξ)|2 dξ.
Rd
Rd Rd
−d
= (2π) 2 e−iz+y,ξ
µ(|z|)U ∗ (z + y, y)ϕ(y)dydz
Rd Rd
= (2π)− 2 e−iy,ξ
ϕ(y)dy e−iz,ξ
µ(|z|)U ∗ (z, 0)dz
d
Rd Rd
%
= ϕ(ξ) e−iz,ξ
µ(|z|)U ∗ (z, 0)dz .
Rd
Since the function µ(|z|)U ∗ (z, 0) depends only on |z|, we can use Lemma 2.13,
i.e. it is sufficient to evaluate the remaining integral in ξ = (0, 0, |ξ|) .
Let us now consider the case d = 3 only, for d = 2 the further steps are
almost the same. Using spherical coordinates,
for r ∈ [0, ∞), φ ∈ [0, 2π), θ ∈ [0, π), we obtain for the remaining integral
1 µ(|z|)
I(|ξ|) = e−iz,ξ
dz
4π |z|
Rd
∞ 2ππ
1 µ(r) 2
= e−i|ξ|r cos θ r sin θdθ dφ dr
4π r
0 0 0
∞ π
1
= r µ(r) e−ir|ξ| cos θ sin θ dθ dr.
2
0 0
π 1 1
−ir|ξ| cos θ 1 −ir|ξ|u 2 sin r|ξ|
e sin θ dθ = e−ir|ξ|u du = − e =
ir|ξ| −1 r|ξ|
0 −1
and therefore
∞
1
I(|ξ|) = µ(r) sin r|ξ| dr .
|ξ|
0
Due to 0 ≤ µ(r) ≤ 1 and since µ(r) has compact support, we further conclude
1
|I(|ξ|)| ≤ c1 (R) for |ξ| ≥ 1.
|ξ|2
Note that
(1 + |ξ|2 )2 ≤ 4 |ξ|4 for |ξ| ≥ 1.
Then,
(1 + |ξ|2 )|%
uµ (ξ)|2 dξ = (1 + |ξ|2 )|I(|ξ|)ϕ(ξ)|
% 2
dξ
|ξ|>1 |ξ|>1
1 + |ξ|2 1
≤ [c1 (R)]2 |ϕ(ξ)|
% 2
dξ ≤ 4[c1 (R)]2 |ϕ(ξ)|
% 2
dξ.
|ξ|4 1 + |ξ|2
|ξ|>1 |ξ|>1
and therefore
|I(|ξ|)| ≤ c2 (R) for |ξ| ≤ 1.
Hence we have
(1 + |ξ|2 )|%
uµ (ξ)|2 dξ = (1 + |ξ|2 )|I(|ξ|)ϕ(ξ)|
% 2
dξ
|ξ|≤1 |ξ|≤1
1
≤ 2 [c2 (R)]2 |ϕ(ξ)|
% 2
dξ ≤ 4[c2 (R)]2 |ϕ(ξ)|
% 2
dξ.
1 + |ξ|2
|ξ|≤1 |ξ|≤1
and therefore
0 ϕH 1 (Ω) ≤ c ϕ -−1 .
N H (Ω)
Hence we have
0 f, ϕΩ |
|N 0 ϕΩ |
|f, N
-−1 (Ω) N0 ϕH 1 (Ω)
f H
= ≤ ≤ c f H
-−1 (Ω)
ϕH
-−1 (Ω) ϕH-−1 (Ω) ϕH-−1 (Ω)
for all ϕ ∈ C0∞ (Ω). When taking the closure with respect to the norm
-−1 (Ω) and using a duality arguments gives (6.3).
· H
Proof. For ϕ ∈ C0∞ (Rd ) we apply integration by parts, exchange the order of
integration, and using the symmetry of the fundamental solution we obtain
[−∆x (N0 f)(x)]ϕ(x)dx = (N 0 f)(x)[−∆x ϕ(x)]dx
Rd Rd
= U ∗ (x, y)f(y)dy[−∆x ϕ(x)]dx
Rd Rd
= f(y) U ∗ (y, x)[−∆x ϕ(x)]dxdy
Rd Rd
= f(y) [−∆x U ∗ (y, x)]ϕ(x)dxdy
Rd Rd
= f(y) δ0 (x − y)ϕ(x)dxdy
Rd Rd
= f(y)ϕ(y)dy.
Rd
When taking the closure of C0∞ (Rd ) with respect to the norm · H 1 (Rd ) this
shows that the partial differential equation (6.4) is satisfied in the sense of
H −1 (Rd ).
116 6 Boundary Integral Operators
satisfying
N0 f H 1/2 (Γ ) ≤ cN
2 f H
-−1 (Ω)
−1 (Ω).
for all f ∈ H (6.6)
and
∗ ∗
lim x, y) − U (x, y)]f (y)dy = 0 .
[U (
x→x∈Γ
Ω
y∈Ω:|y−x|>ε
For the remaining term we obtain
∗
x, y)f (y)dy ≤ f L∞ (Ω∩Bε (x))
U ( |U ∗ (
x, y)|dy
y∈Ω:|y−x|≤ε Ω∩Bε (x)
≤ f L∞ (Ω) |U ∗ (
x, y)|dy.
B2ε (
x)
6.1 Newton Potential 117
2π2ε
1
= | log r| r drdϕ = ε2 [1 − 2 log(2ε)] .
2π
0 0
2ππ 2ε
1 1 2
= r sin ψ drdψdϕ = 2 ε2 .
4π r
0 0 0
−1 (Ω).
is satisfied for all f ∈ H
≤ (c + 1)cIT f H
-−1 (Ω) wH −1/2 (Γ ) .
118 6 Boundary Integral Operators
−∆u(x) = 0 for x ∈ Ω ∪ Ω c .
where
(N0 ϕ)(y) = γ0int U ∗ (x, y)ϕ(x)dx for y ∈ Γ.
Ω
V : H −1/2 (Γ ) → H 1 (Ω).
satisfying
→ x and ε → 0,
The assertion now follows as in the proof of Lemma 6.4 for x
we skip the details.
In the same way we obtain for the exterior trace
satisfying
and
1 1
σ(x) := lim dsy for x ∈ Γ. (6.11)
ε→0 2(d − 1)π εd−1
y∈Ω:|y−x|=ε
x∈Ω:|x−y|=ε
x∈Ω:|x−y|=ε x∈Ω:|x−y|=ε
+ ϕ(y) int U ∗ (x, y)ds
γ1,x x
x∈Ω:|x−y|=ε
with
int U ∗ (x, y)[ϕ(x) − ϕ(y)]ds
γ1,x x
x∈Ω:|x−y|=ε
≤ max |ϕ(x) − ϕ(y)| int U ∗ (x, y)|ds .
|γ1,x x
x∈Ω:|x−y|=ε
x∈Ω:|x−y|=ε
For d = 2 we have
122 6 Boundary Integral Operators
int U ∗ (x, y)|ds ≤
|γ1,x int U ∗ (x, y)|ds
|γ1,x
x x
while for d = 3
int U ∗ (x, y)|ds ≤
|γ1,x int U ∗ (x, y)|ds
|γ1,x
x x
y−x
For the remaining integral we find by using nx = for x ∈ Ω, |y−x| = ε,
|y − x|
int U ∗ (x, y)ds = − 1 (nx , x − y)
γ1,x x dsx
2(d − 1)π |x − y|d
x∈Ω:|x−y|=ε x∈Ω:|x−y|=ε
1 1 1 1
= dsx = dsx .
2(d − 1)π |x − y|d−1 2(d − 1)π εd−1
x∈Ω:|x−y|=ε x∈Ω:|x−y|=ε
Taking into account the definitions (6.10) and (6.11) we finally obtain
γ1int u(x)γ0int ϕ(x)dsx
Γ
⎡ ⎤
⎢ int ∗ ⎥
= w(y) ⎣ lim γ1,x U (x, y)γ0int ϕ(x)dsx + γ0int ϕ(y)σ(y)⎦ dsy
ε→0
Γ x∈Γ :|x−y|≥ε
= γ0int ϕ(x) lim int U ∗ (x, y)w(y)ds ds +
γ1,x w(y)σ(y)ϕ(y)dsy
y x
ε→0
Γ y∈Γ :|y−x|≥ε Γ
= [σ(x)w(x) + (K w)(x)]γ0int ϕ(x)dsx .
Γ
6.3 Adjoint Double Layer Potential 123
Lemma 6.9. For the conormal derivative of the single layer potential V there
holds the jump relation
[γ1 V w] := γ1ext (V w)(x) − γ1int (V w)(x) = −w(x) for x ∈ Γ (6.12)
and therefore
w(x)ϕ(x)dsx = [γ1int u(x) − γ1ext u(x)]γ0int ϕ(x)dsx
Γ Γ
holds for all ϕ ∈ C0∞ (Rd ). The closure of C0∞ (Rd ) with respect to · H 1/2 (Γ )
and a duality argument then gives the assertion.
0 f Γ
W v, f Ω = v, γ1int N
0 f )(x) = f (x)
−∆x (N for x ∈ Ω.
W v, f Ω
W vH 1 (Ω) = sup
-−1 (Ω)
0=f ∈H
f H
-−1 (Ω)
0 f Γ
v, γ1int N
= sup ≤ c vH 1/2 (Γ ) .
-−1 (Ω) f H
0=f ∈H -−1 (Ω)
The double layer potential (6.13) therefore defines a linear and bounded op-
erator
W : H 1/2 (Γ ) → H 1 (Ω).
When applying the interior trace operator γ0int : H 1 (Ω) → H 1/2 (Γ ) to the
double layer potential u = W v ∈ H 1 (Ω) this declares, for v ∈ H 1/2 (Γ ), a
linear and bounded operator
satisfying
where σ(x) is as defined in (6.11) and with the double layer potential
(Kv)(x) := lim int U ∗ (x, y)]v(y)ds
[γ1,y for x ∈ Γ.
y
ε→0
y∈Γ :|y−x|≥ε
y∈Γ :|y−x|≥ε
y∈Γ :|y−x|≥ε
+ int U ∗ (
[γ1,y x, y)]v(y)dsy
y∈Γ :|y−x|<ε
2 3
= int U ∗ (
γ1,y int U ∗ (x, y) v(y)ds
x, y) − γ1,y y
y∈Γ :|y−x|≥ε
+ int U ∗ (
[γ1,y x, y)][v(y) − v(x)]dsy
y∈Γ :|y−x|<ε
+ v(x) int U ∗ (
γ1,y x, y)dsy .
y∈Γ :|y−x|<ε
∈ Ω we further have
For x
int U ∗ (
|γ1,y x, y)|dsy ≤ M .
Γ
Therefore, the second term vanishes when considering the limit ε → 0. For
the computation of the third term we consider
Bε (x) = {y ∈ Ω : |y − x| < ε} .
Then,
γ int U ∗ (
1,y x, y)ds y = int U ∗ (
γ1,y x, y)dsy − int U ∗ (
γ1,y x, y)dsy .
y∈Γ :|y−x|<ε ∂Bε (x) y∈Ω:|y−x|=ε
∈ Bε (x) we
Using the representation formula (6.1) for u = 1 and due to x
obtain
6.4 Double Layer Potential 127
γ1int U ∗ (
x, y)dsy = −1.
∂Bε (x)
Rd Rd Γ
= w(y)γ int
1,y −∆x U ∗ (x, y)ϕ(x)dxdsy
Γ Rd
= int
w(y)γ1,y δ0 (x − y)ϕ(x)dxdsy = 0 .
Γ Rd
For
we first have
DvH −1/2 (Γ ) ≤ cD
2 vH 1/2 (Γ ) for v ∈ H 1/2 (Γ ). (6.16)
However, when taking the limit ε → 0 for x ∈ Γ the integrals does not exist as
Cauchy principal value. As a generalization of the Cauchy integral we there-
fore call D a hypersingular boundary integral operator. To find an explicit
representation of D we therefore have to introduce a suitable regularisation.
Inserting u0 (x) ≡ 1 into the representation formula (6.1) this gives
1 = − γ1,y int U ∗ ( ∈ Ω.
x, y)dsy for x
Γ
Hence we have
∇x (W u0 )( ∈ Ω,
x) = 0 for x
and therefore
(Du0 )(x) = 0 for x ∈ Γ. (6.17)
6.5 Hypersingular Boundary Integral Operator 129
Moreover, /
dsy = [y1 (t)]2 + [y2 (t)]2 dt,
and the exterior normal vector is given by
1 y2 (t)
n(y) = . for y ∈ Γk .
[y1 (t)]2 + [y2 (t)]2 −y1 (t)
v( = x + (
x) = v(x) for x x − x, n(x))n(x).
130 6 Boundary Integral Operators
Γk tk
Theorem 6.15. Let Γ be a piecewise smooth closed curve and let u and v be
globally continuous on Γ . Moreover, let u and v be continuously differentiable
on the parts Γk . Then we can rewrite the bilinear form of the hypersingular
boundary integral operator D as
1
Du, vΓ = − curlΓ v(x) log |x − y| curlΓ u(y)dsy dsx . (6.19)
2π
Γ Γ
we obtain
∂ ∂ ∂
log |
x − y| = −n(y) · ∇y log |
x − y| .
∂
xi ∂ny ∂yi
Due to ∆y log |
x − y| = 0 for y = x we further get
∂
curlΓ,y log |
x − y| =
∂y1
∂ ∂ ∂ ∂
= n1 (y) log |
x − y| − n2 (y) log |
x − y|
∂y2 ∂y1 ∂y1 ∂y1
∂ ∂ ∂ ∂
= n1 (y) log |
x − y| + n2 (y) log |
x − y|
∂y2 ∂y1 ∂y2 ∂y2
∂
= n(y) · ∇y log |
x − y| ,
∂y2
and
∂ ∂
curlΓ,y log |
x − y| = −n(y) · ∇y log |
x − y| .
∂y2 ∂y1
Hence we can write the partial derivatives of the double layer potential for a
globally continuous function u by applying integration by parts as
132 6 Boundary Integral Operators
∂ 1 ∂ ∂
x) = −
w( u(y) log |
x − y| dsy
∂
x1 2π ∂
x1 ∂ny
Γ
1 ∂
= u(y)n(y) · ∇y log |
x − y| dsy
2π ∂y1
Γ
1 ∂
=− u(y) curlΓ,y log |x − y| dsy
2π ∂y2
Γ
1 ∂
= curlΓ u(y) log |
x − y| dsy ,
2π ∂y2
Γ
and
∂ 1 ∂
x) = −
w( curlΓ u(y) log |
x − y| dsy .
∂
x2 2π ∂y1
Γ
For the normal derivative of the double layer potential we then obtain
Therefore,
6.5 Hypersingular Boundary Integral Operator 133
∂ 1
v(x) w(x)dsx = − v(x) lim curlΓ u(y) curlΓ,x log |x − y| dsy dsx
∂nx 2π ε→0
Γ x∈Γ y∈Γ :|y−x|≥ε
1
=− curlΓ u(y) lim v(x) curlΓ,x log |x − y| dsx dsy
2π ε→0
y∈Γ x∈Γ :|x−y|≥ε
1
= curlΓ u(y) lim curlΓ v(x) log |x − y| dsx dsy ,
2π ε→0
y∈Γ x∈Γ :|x−y|≥ε
∇ × [
u(x)v(x)] = ∇
u(x) × v(x) + u
(x) [∇ × v(x)]
we obtain
curlΓk u(x) · v(x)dsx = [n(x) × ∇
u(x)] · v(x)dsx
Γk Γk
= u(x) × v(x)] · n(x) dsx
[∇
Γk
= [∇ × [
u(x)v(x)] − u
(x) [∇ × v(x)]] · n(x) dsx
Γk
= u(x)v(x)t(x)dσ − u(x) curlΓk v(x)dsx
∂Γk Γk
Theorem 6.17. Let Γ be a piecewise smooth closed surface, and let u and v
be globally continuous functions defined on Γ which are differentiable on Γk .
Then the bilinear form of the hypersingular boundary integral operator D can
be written as
1 curlΓ u(y) · curlΓ v(x)
Du, vΓ = dsx dsy .
4π |x − y|
Γ Γ
∈Ω
Proof. The proof follows essentially as in the two–dimensional case. For x
and using the definition (6.15) of the hypersingular boundary integral operator
D we have to consider the double layer potential
1 ∂ 1
w(x) := − u(y) dsy .
4π ∂ny |
x − y|
Γ
Using
∂ 1 i − yi
x yi − x
i ∂ 1
= = − = −
∂yi |
x − y| x − y|3
| |
x − y| xi |
∂ x − y|
we obtain for the partial derivatives of the kernel function
∂ ∂ 1 ∂ 1
= −n(y) · ∇y .
xi ∂ny |
∂ x − y| ∂yi |
x − y|
and hence we can write the gradient of the double layer potential as
1 1
∇x w(
x) = − curlΓ,y u(y) × ∇y dsy
4π |
x − y|
Γ
1 1
= curlΓ,y u(y) × ∇x dsy .
4π |
x − y|
Γ
By using
0 1
curlΓ,x [v(x)curlΓ,y u(y)] = n(x) · ∇x × [v(x)curlΓ,y u(y)]
0 1
= n(x) · ∇x v(x) × curlΓ,y u(y)
= [n(x) × ∇x v(x)] · curlΓ,y u(y)
= curlΓ,x v(x) · curlΓ,y u(y)
we finally conclude the assertion.
γ0int u(x) = (V γ1int u)(x) + [1 − σ(x)]γ0int u(x) − (Kγ0int u)(x) + N0 f (x). (6.20)
γ1int u(x) = σ(x)γ1int u(x) + (K γ1int u)(x) + (Dγ0int u)(x) + N1 f (x). (6.21)
With (6.20) and (6.21) we have obtained a system of two boundary integral
equations which can be written for x ∈ Γ as
& ' & '& ' & '
γ0int u (1 − σ)I − K V γ0int u N0 f
= + (6.22)
γ1int u D σI + K γ1int u N1 f
Proof. Let (ψ, ϕ) ∈ H −1/2 (Γ ) × H 1/2 (Γ ) be arbitrary but fixed. The function
x) := (V ψ)(
u( x) − (W ϕ)(
x) ∈Ω
for x
This is equivalent to
& ' & '& '
γ0int u(x) (1 − σ)I − K V γ0int u(x)
= .
γ1int u(x) D σI + K γ1int u(x)
Inserting
138 6 Boundary Integral Operators
& ' & '& '
γ0int u(x) (1 − σ)I − K V ϕ(x)
=
γ1int u(x) D σI + K ψ(x)
V K = KV, (6.26)
K D = DK. (6.27)
Note that (6.26) describes the symmetrization of the double layer potential
K, which is in general not self–adjoint, by the single layer potential V . This
property was already described by J. Plemelj in 1911 in the case of the two–
dimensional Laplace operator [112].
From the system (6.22) of boundary integral equations we may also find
a suitable representation of the Newton potential N1 f when assuming the
invertibility of the single layer potential V , see also Subsection 6.6.1.
Lemma 6.20. For the volume potential (N1 f )(x), x ∈ Γ , there holds the
representation
Proof. Using the first boundary integral equation in (6.22) and assuming the
invertibility of the single layer potential V we first obtain
Inserting this into the second boundary integral equation of (6.22) we get
cint int 2
1 γ1 uH −1/2 (Γ ) ≤ aΩ (u, u) . (6.29)
and
1
|∇u(x)| = |∇(V w)(x)| ≤ c2 (w) . (6.32)
|x − y0 |2
Proof. By using the triangle inequality we have for y ∈ Ω
1
|x − y0 | ≤ |x − y| + |y − y0 | ≤ |x − y| + diam(Ω) ≤ |x − y| + |x − y0 |
2
140 6 Boundary Integral Operators
and therefore
1
|x − y0 | .
|x − y| ≥
2
In the case d = 3 we find the estimate (6.31) from
|u(x)| = γ0int U ∗ (x, ·), wΓ
≤ γ0int U ∗ (x, ·)H 1/2 (Γ ) wH −1/2 (Γ ) ≤ cT ||U ∗ (x, ·)||H 1 (Ω) ||w||H −1/2 (Γ )
and by using
1 1 1 1
||U ∗ (x, ·)||2H 1 (Ω) = dy + dy
16π 2 |x − y|2 16π 2 |x − y|4
Ω Ω
1 1 1 1 5 |Ω| 1
≤ 2 2
dy+ 2 4
dy ≤ .
4π |x − y0 | π |x − y0 | 4 π |x − y0 |2
2
Ω Ω
Then, u(x) = (V w)(x) for x ∈ Ω c is the unique solution of the Dirichlet
boundary value problem
Using Green’s first formula with respect to the bounded domain BR (y0 )\Ω
this gives
6.6 Properties of Boundary Integral Operators 141
aBR (y0 )\Ω (u, v) = −γ1ext u, γ0ext vΓ + γ1int u, γ0int v∂BR (y0 )
where we have used the opposite direction of the exterior normal vector along
Γ . Choosing v = u and using Lemma 6.21 we have
int 1
γ1 u, γ0int u∂BR (y0 ) ≤ c1 (w)c2 (w) dsx ≤ c Rd−4 .
|x − y0 |3
|x−y0 |=R
Hence we can consider the limit R → ∞ to obtain Green’s first formula for
u = V w with respect to the exterior domain as
aΩ c (u, u) := ∇u(x)∇u(x)dx = −γ1ext u, γ0ext uΓ . (6.33)
Ωc
−1/2
Note that in the two–dimensional case the assumption w ∈ H∗ (Γ ) is es-
sential to ensure the above result. In analogy to the estimate (4.17) for the
solution of the interior Dirichlet boundary value problem we find
cext
1 γ1ext u2H −1/2 (Γ ) ≤ aΩ c (u, u). (6.34)
−1/2
Theorem 6.22. Let w ∈ H −1/2 (Γ ) for d = 3 and w ∈ H∗ (Γ ) for d = 2,
respectively. Then there holds
Proof. For u = V w we can apply both the interior and exterior Green’s for-
mulae, i.e. (6.28) and (6.33) to obtain
Taking the sum of the above equations we obtain from the jump relation (6.9)
of the single layer potential
The jump relation (6.12) of the conormal derivate of the single layer potential
reads
γ1int u(x) − γ1ext u(x) = w(x) for x ∈ Γ
and therefore we have
−1/2
In the two–dimensional case we only have the H∗ (Γ ) ellipticity of the single
layer potential when using the previous theorem. To obtain a more general
result we first consider the following saddle point problem, d = 2, 3, to find
(t, λ) ∈ H −1/2 (Γ ) × R such that
λ = V weq , weq Γ .
capΓ := e−2πλ
satisfying
1 1 r
(Vr weq )(x) = log r + λ = log .
2π 2π capΓ
In particular for r = 1 we obtain
1 1
λ := log .
2π capΓ
If the logarithmic capacity capΓ < 1 is strictly less than one, we conclude λ >
0. To ensure capΓ < 1 a sufficient criteria is to assume diam Ω < 1 [81, 157].
This assumption can be always guaranteed when considering a suitable scaling
of the domain Ω ⊂ R2 .
Theorem 6.23. For d = 2 let diam(Ω) < 1 and therefore λ > 0 be satisfied.
The single layer potential V is then H −1/2 (Γ )–elliptic, i.e.,
cV1 w2H −1/2 (Γ )
V w, wΓ ≥ for all w ∈ H −1/2 (Γ ).
satisfying
V w, wΓ = V (w
+ αweq ), w
+ αweq Γ
= V w,
w Γ + α2 V weq , weq Γ
Γ + 2α V weq , w
2H −1/2 (Γ ) + α2 λ
≥ cV1 w
2 3
≥ min{cV1 , λ} w 2H −1/2 (Γ ) + α2 ,
−1/2 1/2
Thus, V : H∗ (Γ ) → H∗ (Γ ) is an isomorphism.
Due to (6.17) we have (Du0 )(x) = 0 with the eigensolution u0 (x) ≡ 1 for
x ∈ Γ . Hence we can not ensure the ellipticity of the hypersingular bound-
ary integral operator D on H 1/2 (Γ ). Instead we have to consider a suitable
subspace.
1/2
Theorem 6.24. The hypersingular boundary integral operator D is H∗ (Γ )–
elliptic, i.e.,
2 1/2
Dv, vΓ ≥ cD
1 vH 1/2 (Γ ) for all v ∈ H∗ (Γ ).
1/2
Proof. For v ∈ H∗ (Γ ) we consider the double layer potential
and
By taking the sum of both Green’s formulae with respect to the interior and
to the exterior domain, and considering the jump relations of the boundary in-
tegral operators involved, we obtain for the bilinear form of the hypersingular
boundary integral operator
Dv, vΓ = γ1int u, [γ0int u − γ0ext u]Γ = γ1int u, γ0int uΓ − γ1ext u, γ0ext uΓ
= |∇u(x)|2 dx + |∇u(x)|2 dx = |u|2H 1 (Ω) + |u|2H 1 (Ω c ) .
Ω Ωc
For the exterior domain Ω c we find from the far field behavior of the double
layer potential u(x) = −(W v)(x) as |x| → ∞ the norm equivalence
c1 u2H 1 (Ω c ) ≤ |u|2H 1 (Ω c ) ≤ c2 u2H 1 (Ω c ) .
1/2
For v ∈ H∗ (Γ ), for the natural density weq ∈ H −1/2 (Γ ), V weq = 1, and by
using the symmetry relation (6.26) we further obtain
1 1
γ0int u, weq Γ = ( I − K)v, weq Γ = v, weq Γ − ( I + K)v, weq Γ
2 2
1 −1 −1 1
= −( I + K)v, V 1Γ = −V ( I + K)v, 1Γ
2 2
1 −1 −1 1
= −( I + K )V v, 1Γ = −V v, ( I + K)1Γ = 0
2 2
1/2
and therefore γ0int u ∈ H∗ (Γ ). By using the norm equivalence theorem of
Sobolev (Theorem 2.6) we find
! "1/2
uH∗1 (Ω) := [γ0int u, weq Γ ]2 + ∇u2L2 (Ω)
1/2 1/2
to be an equivalent norm in H 1 (Ω). For v ∈ H∗ (Γ ) we have γ0int u ∈ H∗ (Γ )
and therefore
By using the trace theorem and the jump relation of the double layer potential
we obtain
! "
Dv, vΓ ≥ c u2H 1 (Ω) + u2H 1 (Ω c )
! "
≥ c γ0int u2H 1/2 (Γ ) + γ0ext u2H 1/2 (Γ )
1
≥ c γ0int u − γ0ext u2H 1/2 (Γ ) = cD
2
1 vH 1/2 (Γ )
2
1/2 1/2
for all v ∈ H∗ (Γ ) and therefore the H∗ (Γ )–ellipticity of the hypersingular
boundary integral operator D.
6.6 Properties of Boundary Integral Operators 147
1/2
The definition of H∗ (Γ ) involves the natural density weq ∈ H −1/2 (Γ ) as
the unique solution of the boundary integral equation (6.36). From a practi-
cal point of view, this seems not to be very convenient for a computational
realization. Hence we may use a subspace which is induced by a much simpler
inner product. For this we define
! "
1/2
H∗∗ (Γ ) := v ∈ H 1/2 (Γ ) : v, 1Γ = 0 .
1/2
From (6.37) we then have for v ∈ H∗∗ (Γ )
2
Dv, vΓ ≥ c̄D
1 |v|H 1/2 (Γ )
! "
= c̄D
1 |v|2H 1/2 (Γ ) + [v, 1Γ ]2 ≥
cD 2
1 vH 1/2 (Γ ) (6.38)
1/2
the H∗∗ (Γ )–ellipticity of D where we again used the norm equivalence the-
orem of Sobolev (cf. Theorem 2.6).
1/2 (Γ0 ) let
We finally consider an open surface Γ0 ⊂ Γ . For a given v ∈ H
1/2
v ∈ H (Γ ) denote the extension defined by
v(x) for x ∈ Γ0 ,
v(x) =
0 elsewhere.
1/2 (Γ0 )
to be an equivalent norm in H 1/2 (Γ ). Hence we have for v ∈ H
148 6 Boundary Integral Operators
2 3
Dv, vΓ0 = D
v , vΓ ≥ c̄D v |2H 1/2 (Γ ) = c̄D
1 | 1 v 2L2 (Γ \Γ0 ) + |
v |2H 1/2 (Γ )
Since the single layer potential V is invertible, we get from the first boundary
integral equation a representation for the Dirichlet to Neumann map,
The operator
With (6.40) and (6.42) we have described the Dirichlet to Neumann map
which maps some given Dirichlet datum γ0int u ∈ H 1/2 (Γ ) to the corresponding
Neumann datum γ1int u ∈ H −1/2 (Γ ) of the harmonic function u ∈ H 1 (Ω)
satisfying Lu = 0.
By using the H 1/2 (Γ )–ellipticity of the inverse single layer potential V −1
we obtain
6.6 Properties of Boundary Integral Operators 149
Sv, vΓ = Dv, vΓ + V −1 (σI + K)v, (σI + K)vΓ ≥ Dv, vΓ (6.45)
as well as
2 1/2
Sv, vΓ ≥ 1 vH 1/2 (Γ )
cD for all v ∈ H∗∗ (Γ ) (6.47)
while for Γ0 ⊂ Γ we get
with 8
1 1
cK = + − cV1 cD
1 <1 (6.50)
2 4
where cV1 and cD
1 are the ellipticity constants of the single layer potential V
and of the hypersingular boundary integral operator D, respectively.
and with
it follows that
Su, uΓ ≤ (σI + K)uV −1 uV −1 .
1/2
Since the hypersingular integral operator D is elliptic for u ∈ H∗ (Γ ) we
find from the mapping properties of the inverse single layer potential V −1 the
lower estimate
2 −1 2
Du, uΓ ≥ cD
1 uH 1/2 (Γ ) ≥ c1 c1 V
D V
1 c1 uV −1 .
u, uΓ = cD V
Denoting
a := (σI + K)uV −1 ≥ 0, b := uV −1 > 0
we conclude 9 a :2 a
− 1 ≤ 0
+ cV1 cD
b b
which is equivalent to
8 8
1 1 a 1 1
− − c1 c1 ≤
V D ≤ + − cV1 cD
1
2 4 b 2 4
and therefore to the assertion.
The contraction property of σI + K, in particular the upper estimate in
(6.49), can be extended to hold in H 1/2 (Γ ).
6.6 Properties of Boundary Integral Operators 151
[u, weq Γ ]2
u2V −1 =
u2V −1 + u2V −1
≥
1, weq Γ
the lower estimate in (6.52). Moreover, using both representations (6.41) and
(6.43) of the Steklov–Poincaré operator S, we conclude
(1 − cK ) wV ≤ (σI + K )wV ≤ cK wV (6.53)
where the contraction rate cK < 1 is given as in (6.50) and · V is the norm
which is induced by the single layer potential V .
−1/2 1/2
Proof. For w ∈ H∗ (Γ ) there exists a uniquely determined v ∈ H∗ (Γ )
satisfying v = V w or w = V −1 v. Using the symmetry property (6.26) we first
have
as well as
w2V = V w, wΓ = V −1 v, vΓ = v2V −1 .
Therefore, (6.53) is equivalent to (6.49).
As in Corollary 6.27 we can extend the contraction property of σI + K
to H −1/2 (Γ ).
(σI + K )wV ≤ cK wV . (6.54)
0 : H
Theorem 6.32. The Newton potential N s (Ω) → H s+2 (Ω) is a continu-
ous map for all s ∈ [−2, 0], i.e.
0 f H s+2 (Ω) ≤ c f -s
N s (Ω).
for all f ∈ H
H (Ω)
0 f, gΩ
N 0 gΩ
f, N
0 f H s+2 (Ω) =
N sup = sup
-−2−s (Ω) gH
0=g∈H -−2−s (Ω) -−2−s (Ω) gH
0=g∈H -−2−s (Ω)
0 gH −s (Ω)
N
≤ f H
-s (Ω) sup ≤ c f H
-s (Ω) .
-−2−s (Ω) gH
0=g∈H -−2−s (Ω)
V wH 1/2+s (Γ ) ≤ c wH −1/2+s (Γ )
V : H −1/2+s (Γ ) → H 1/2+s (Γ ),
K : H 1/2+s (Γ ) → H 1/2+s (Γ ),
K : H −1/2+s (Γ ) → H −1/2+s (Γ ),
D : H 1/2+s (Γ ) → H −1/2+s (Γ )
Proof. For the single layer potential V and for |s| < 12 the assertion was
already shown in Theorem 6.33. This remains true for |s| = 12 [152], see also
the discussion in [103].
For all other boundary integral operators the assertion follows from the
mapping properties of the conormal derivative operator.
First we consider the adjoint double layer potential K . Recall that the sin-
gle layer potential u(x) = (V w)(x), x ∈ Ω, is a solution of the homoge-
neous partial differential equation with Dirichlet data γ0int u(x) = (V w)(x) for
x ∈ Γ . By using Theorem 4.6 and the continuity of the single layer potential
V : L2 (Γ ) → H 1 (Γ ) we obtain
γ1int uL2 (Γ ) ≤ c V wH 1 (Γ ) ≤
c wL2 (Γ )
γ1int u, ϕΓ
γ1int uH −1 (Γ ) = sup .
0=ϕ∈H 1 (Γ ) ϕH 1 (Γ )
γ1int vL2 (Γ ) ≤ c γ0int vH 1 (Γ ) = c ϕH 1 (Γ ) .
v, K wΓ
= sup
0=w∈H −1/2−s (Γ ) wH −1/2−s (Γ )
K wH −1/2−s (Γ )
= vH 1/2+s (Γ ) sup
0=w∈H −1/2−s (Γ ) wH −1/2−s (Γ )
≤ c vH 1/2+s (Γ )
for i, j = 1, . . . , d where
⎧
⎨ − log |x − y|
⎪ for d = 2,
E(x, y) = 1
⎪
⎩ for d = 3.
|x − y|
For the components ui of the solution there holds the representation formula
∈ Ω, i = 1, . . . , d,
(5.10) (Somigliana identity), x
d
d
∗
ui (
x) = Uij (
x, y)tj (y)dsy − Tij∗ (
x, y)uj (y)dsy
Γ j=1 Γ j=1
d
∗
+ Uij (
x, y)fj (y)dy. (6.57)
Ω j=1
0 f [H 1 (Ω)]d ≤ c f -−1
N .
[H (Ω)]d
0 f )i (x) =
(N0 f )i (x) := γ0int (N lim 0 f )i (
(N x) for i = 1, . . . , d,
x→x∈Γ
Ω
0 : [H
N0 := γ0int N −1 (Ω)]d → [H 1/2 (Γ )]d .
d
0 f )i (x) =
(γ1int N lim 0 f , x
σij (N )nj (x) for i = 1, . . . , d,
x→x∈Γ
Ω
j=1
d
(γ1int V w)i (x) = lim σij (V w, x
)nj (x) for i = 1, . . . , d,
x→x∈Γ
Ω
j=1
Proof. The ellipticity estimate follows as is the proof of Theorem 6.22 by using
Lemma 4.19.
To prove the [H −1/2 (Γ )]2 –ellipticity of the two–dimensional single layer
potential V we first introduce the generalized fundamental solution
1 1 1+ν (xi − yi )(xj − yj )
α
Uij (x, y) = (4ν − 3) log(α|x − y|)δij +
4π E 1 − ν |x − y|2
for i, j = 1, 2 which depends on a real parameter α ∈ R+ , and we consider the
corresponding single layer potential Vα : [H −1/2 (Γ )]2 → [H −1/2 (Γ )]2 . Note
that this approach corresponds to some scaling of the computational domain
Ω ⊂ R2 and its boundary Γ , respectively.
−1/2
For w ∈ [H+ (Γ )]2 we have by using Theorem 6.36 the ellipticity esti-
mate
Vα w, wΓ = V w, wΓ ≥ cV1 w2[H −1/2 (Γ )]2 .
The further approach now corresponds to the case of the scalar single layer
potential of the Laplace operator [142] to find (w1 , λ1 ) ∈ [H −1/2 (Γ )]2 × R2 as
the solution of the saddle point problem
Vα w1 , τ Γ − λ11 1, τ1 Γ − λ12 1, τ2 Γ = 0
w11 , 1Γ =1
1
w2 , 1Γ =0
λ22 = Vα w2 , w2 Γ ,
as well as
λ12 = λ21 = Vα w1 , w2 Γ .
160 6 Boundary Integral Operators
Lemma 6.37. For the Lagrange multiplier λ1i (i = 1, 2) we have the repre-
sentation
1 1 1+ν
λii = V wi , wi Γ + (4ν − 3) log α,
4π E 1 − ν
while the Lagrange multiplier λ12 = λ21 is independent of α ∈ R+ ,
λ12 = λ21 = V w1 , w2 Γ .
Proof. For i = 1, a direct computation gives, by splitting the fundamental
solution log(α|x − y|),
λ11 = Vα w1 , w1 Γ
2
1 1 1+ν
= (4ν − 3) log(α|x − y|)wi1 (y)wi1 (x)dsx dsy
4π E 1 − ν i=1
Γ Γ
2
1 1 1+ν (xi − yi )(xj − yj ) 1
+ 2
wi (y)wj1 (x)dsx dsy
4π E 1 − ν i,j=1
|x − y|
Γ Γ
2
1 1 1+ν
= (4ν − 3) log |x − y|wi1 (y)wi1 (x)dsx dsy
4π E 1 − ν i=1Γ Γ
2
1 1 1+ν (xi − yi )(xj − yj ) 1
+ 2
wi (y)wj1 (x)dsx dsy
4π E 1 − ν i,j=1
|x − y|
Γ Γ
2
1 1 1+ν 0 1 12
+ (4ν − 3) log α wi , 1Γ
4π E 1 − ν i=1
1 1 1+ν
= V1 w1 , w1 Γ + (4ν − 3) log α
4π E 1 − ν
due to w11 , 1Γ = 1 and w21 , 1Γ = 0. For λ22 the assertion follows in the same
way. Finally, for λ21 = λ12 we have
λ21 = Vα w1 , w2 Γ
2
1 1 1+ν
= V1 w1 , w2 Γ + (4ν − 3) log α wi1 , 1Γ wi2 , 1Γ
4π E 1 − ν i=1
= V1 w1 , w2 Γ
due to w12 , 1Γ = w21 , 1Γ = 0.
Hence we can choose the scaling parameter α ∈ R+ such that
min{λ11 , λ22 } ≥ 2 |λ12 | (6.59)
is satisfied. An arbitrary given w ∈ [H −1/2 (Γ )]2 can be written as
+ α1 w1 + α2 w2 ,
w = w αi = wi , 1Γ (i = 1, 2) (6.60)
−1/2
∈ [H+
where w (Γ )]2 .
6.7 Linear Elasticity 161
Theorem 6.38. Let the scaling parameter α ∈ R+ be chosen such that (6.59)
is satisfied. Then the single layer potential Vα is [H −1/2 (Γ )]2 –elliptic, i.e.
cV1 w2[H −1/2 (Γ )]2
Vα w, wΓ ≥ for all w ∈ [H −1/2 (Γ )]2 .
and
! "
1/2
[H∗ (Γ )]d := v ∈ [H 1/2 (Γ )]d : V −1 v, v k Γ = 0 for v k ∈ R .
−1/2 1/2
Obviously, V : [H∗ (Γ )]d → [H∗ (Γ )]d is an isomorphism.
∈ Ω ∪ Ω we define by
For x c
d
(W v)i (
x) := Tij∗ (
x, y)uj (y)dsy , i = 1, . . . , d,
Γ j=1
The application of the interior trace operator defines a bounded linear oper-
ator
γ0int W : [H 1/2 (Γ )]d → [H 1/2 (Γ )]d
with the representation
1
(γ0int W v)i (x) = − vi (x) + (Kv)i (x) for almost all x ∈ Γ, i = 1, . . . , d,
2
(6.61)
where
d
(Kv)i (x) := lim Tij∗ (
x, y)uj (y)dsy , i = 1, . . . , d
ε→0
j=1
y∈Γ :|y−x|≥ε
Inserting the rigid body motions (1.36) for d = 2 and (1.29) for d = 3 this
gives
1
I + K v k (x) = 0 for x ∈ Γ and v k ∈ R.
2
6.7 Linear Elasticity 163
The application of the interior boundary stress operator γ1int on the double
layer potential W v defines a bounded linear operator
The bilinear form of the hypersingular boundary integral operator D can then
be written as [75]
& 3 '
µ 1 ∂ ∂
Du, vΓ = u(y) · v(x) dsy dsx
4π |x − y| ∂Sk (y) ∂Sk (x)
Γ Γ k=1
µ I 2 ∗
+ (M (∂x , n(x))v(x)) −4µ U (x, y) M (∂y , n(y))u(y)dsy dsx
2π |x − y|
ΓΓ
3
µ 1
+ Mkj (∂x , n(x))vi (x) Mki (∂y , n(y))vj (y)dsy dsx . (6.64)
4π |x − y|
Γ Γ i,j,k=1
Hence we can express the bilinear form of the hypersingular boundary integral
operator D by components of the single layer potential V only.
Moreover, for d = 3 there holds a related representation of the double layer
potential K, see [89],
1 ∂ 1 1 1
(Ku)(x) = u(y)dsy − M (∂y , n(y))u(y)dsy
4π ∂n(y) |x − y| 4π |x − y|
Γ Γ
E
+ (V (M (∂· , n(·))u(·)))(x)
1+ν
where the evaluation of the Laplace single and double layer potentials has to
be taken componentwise.
Hence we can reduce all boundary integral operators of linear elastostatics
to the single and double layer potentials of the Laplace equation. These rela-
tions can be used when considering the Galerkin discretization of boundary
integral equations, where only weakly singular surface integrals have to be
computed.
Inserting the rigid body motions (1.36) for d = 2 and (1.29) for d = 3 into
the representation formula (6.57) this gives
x) = −(W v k )(
v k ( x) ∈ Ω and v k ∈ R.
for x
of functions which are orthogonal to the rigid body motions. Then we have,
as in (6.38),
2 1/2
Dv, vΓ ≥ 1 v[H 1/2 (Γ )]d
cD for all v ∈ [H∗∗ (Γ )]d .
Since the fundamental solution of the Stokes system coincides with the Kelvin
fundamental solution of linear elastostatics when considering ν = 12 and E = 3
as material parameters, we can write the representation formula (6.57) and
all related boundary integral operators of linear elastostatics for ν = 12 and
E = 3 to obtain the Stokes case. However, since the analysis of the mapping
properties of all boundary integral operators of linear elastostatics assumes
ν ∈ (0, 12 ) we can not transfer the boundedness and ellipticity estimates from
linear elastostatics to the Stokes system. These results will be shown by con-
sidering the Stokes single layer potential which is also of interest for the case
of almost incompressible linear elasticity (ν = 12 ) [141].
Let Ω ⊂ Rd be a simple connected domain with boundary Γ = ∂Ω. The
single layer potential V : [H −1/2 (Γ )]d → [H 1 (Ω)]d induces a function
d
x) := (V w)i (
ui ( x) = ∗
Uij (
x, y)wj (y)dsy ∈ Ω, i = 1, . . . , d
for x
Γ j=1
for all v ∈ [H 1 (Ω)]d with div v = 0. The application of the interior trace
operator defines the single layer potential
166 6 Boundary Integral Operators
int/ext
V := γ0 V : [H −1/2 (Γ )]d → [H 1/2 (Γ )]d
Hence we can expect the ellipticity of the Stokes single layer potential V only
in a subspace which is orthogonal to the exterior normal vector n.
Let V L : H −1/2 (Γ ) → H 1/2 (Γ ) be the Laplace single layer potential which
is H −1/2 (Γ )–elliptic. Hence we can define
d
w, τ V L := V L wi , τi Γ
i=1
As for the homogeneous Neumann boundary value problem for the Laplace
equation we can introduce an extended bilinear form
where the fundamental solution is, see (5.20) for d = 2 and (5.18) for d = 3,
⎧
⎪
⎪
1
⎨ Y0 (k|x − y|) for d = 2,
∗ 2π
Uk (x, y) =
⎪
⎪ 1 eik|x−y|
⎩ for d = 3.
4π |x − y|
Then we can define the standard boundary integral operators for x ∈ Γ , i.e.
the single layer potential
(Vk w)(x) = Uk∗ (x, y)w(y)dsy ,
Γ
As for the Laplace operator there hold all the mapping properties as given in
Theorem 6.34. In particular, Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded, but not
H −1/2 (Γ )–elliptic. However, the single layer potential is coercive satisfying a
Gårdings inequality.
For x ∈ Ω we consider the function
u(x) = (Vk w)(x) − (V w)(x) = [Uk∗ (x, y) − U ∗ (x, y)]w(y)dsy (6.66)
Ω
∗
where U (x, y) is the fundamental solution of the Laplace operator. In par-
ticular we have for y ∈ Γ and x ∈ Ω
Moreover,
−∆x [−∆x − k 2 ]u(x) = −k 2 ∆x U ∗ (x, y)w(y)dsy = 0,
Γ
i.e. the function u as defined in (6.66) solves the partial differential equation
which is of fourth order. Hence, we obtain as in the case of the Laplace oper-
ator, by considering the corresponding Newton potentials, that
Thus,
Vk − V = γ0int [Vk − V ] : H −1/2 (Γ ) → H 5/2 (Γ ),
and by the compact imbedding of H 5/2 (Γ ) in H 1/2 (Γ ) we conclude that
is compact.
C = V − Vk : H −1/2 (Γ ) → H 1/2 (Γ )
we have
Dk − D : H 1/2 (Γ ) → H −1/2 (Γ ),
Kk − K : H 1/2 (Γ ) → H 1/2 (Γ ),
KK − K : H −1/2 (Γ ) → H −1/2 (Γ )
are compact where D, K, K are the hypersingular integral operator, the dou-
ble layer potential and its adjoint of the Laplace operator, respectively.
As for the Laplace operator the bilinear form of the hypersingular bound-
ary integral operator Dk can be written as, by using integration by parts
[107],
ı k|x−y|
1 e
Dk u, vΓ = curlΓ u(y), curlΓ v(x) dsy dsx
4π |x − y|
Γ Γ
k2 eı k|x−y|
− u(y)v(x) n(x), n(y) dsy dsx . (6.68)
4π |x − y|
Γ Γ
6.10 Exercises
Let Γ = ∂Ω be the boundary of the circle Ω = Br (0) ⊂ R2 which can be
described by using polar coordinates as
& '
cos 2πt
x(t) = r ∈ Γ for t ∈ [0, 1).
sin 2πt
conditions such that the single layer potential is invertible, and positive
definite.
6.5 Prove Corollary 6.19.
6.6 Determine the eigenfunctions of the hypersingulur boundary integral op-
erator D for x ∈ ∂Br (0) and compute the corresponding eigenvalues.
6.7 Let now Γ be the boundary of an ellipse given by the parametrization
& '
a cos 2πt
x(t) = ∈ Γ for t ∈ [0, 2π).
b sin 2πt
6.8 Prove that the eigenfunctions of the double layer potential as considered
in Exercise 6.7 are given by
⎧
⎪
⎨ cos 2πkt for k > 0,
vk (t) = 1 for k = 0,
⎪
⎩
sin 2πkt for k < 0.
Hence we have to find the complete Cauchy data γ0int u(x) and γ1int u(x) for
x ∈ Γ , which are given by boundary conditions only partially. For this we
will describe appropriate boundary integral equations. The starting point is
the representation formula (7.2) and the related system (6.22) of boundary
integral equations,
& ' & '& '
γ0int u [1 − σ]I − K V γ0int u
= . (7.3)
γ1int u D σI + K γ1int u
This approach is called direct where the density functions of all boundary
integral operators are just the Cauchy data [γ0int u(x), γ1int u(x)], x ∈ Γ . When
172 7 Boundary Integral Equations
When using the direct approach (7.2) we obtain the representation formula
u(
x) = U ∗ ( int U ∗ (
x, y)γ1int u(y)dsy − γ1,y ∈ Ω (7.7)
x, y)g(y)dsy for x
Γ Γ
where we have to find the yet unknown Neumann datum γ1int u ∈ H −1/2 (Γ ).
By using the first boundary integral equation in (7.3) we obtain with
a first kind Fredholm boundary integral equation. Since the single layer poten-
tial V : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded (see (6.8)) and H −1/2 (Γ )–elliptic
(see Theorem 6.22 for d = 3 and Theorem 6.23 for d = 2 when assuming
diam(Ω) < 1), we conclude the unique solvability of the boundary integral
equation (7.8) when applying the Lax–Milgram lemma (Theorem 3.4). More-
over, the unique solution γ1int u ∈ H −1/2 (Γ ) satisfies
1 cW
2
γ1int uH −1/2 (Γ ) ≤ (σI + K)gH 1/2 (Γ ) ≤ gH 1/2 (Γ ) .
cV1 cV1
7.1 Dirichlet Boundary Value Problem 173
Since the boundary integral equation (7.8) is formulated in H 1/2 (Γ ), this gives
The convergence of the series (7.11) in H −1/2 (Γ ) follows from the contraction
property (6.54) of σI +K when considering the equivalent Sobolev norm ·V
which is induced by the single layer potential V .
When using the indirect single layer potential ansatz (7.4) to find the un-
known density w ∈ H −1/2 (Γ ) we have to solve the boundary integral equation
Note that the boundary integral equation (7.12) differs from the boundary
integral equation (7.8) of the direct approach only in the definition of the
right hand side. Hence we can conclude the unique solvability of the boundary
integral equation (7.12) as for (7.8).
By using the double layer potential (7.5) to describe the solution of the
homogeneous partial differential equation we obtain from the jump relation
(6.14) of the double layer potential the boundary integral equation
The convergence of the series (7.14) in H 1/2 (Γ ) follows from the contraction
property (6.51) of σI + K when considering the equivalent Sobolev norm
· V −1 which is induced by the inverse single layer potential V −1 .
To obtain a variational formulation of the boundary integral equation
(7.13) in H 1/2 (Γ ) we first consider
12 v − Kv − g, τ Γ
0 = [1 − σ]v − Kv − gH 1/2 (Γ ) = sup
0=τ ∈H −1/2 (Γ ) τ H −1/2 (Γ )
( 12 I − K)v, τ Γ
cS vH 1/2 (Γ ) ≤ sup for all v ∈ H 1/2 (Γ )
0=τ ∈H −1/2 (Γ ) τ H −1/2 (Γ )
By using the contraction estimate (6.51) and the mapping properties of the
single layer potential V we obtain
1 1
( I − K)v, τv Γ = ( I − K)v, V −1 vΓ
2 2
1
= V −1 v, vΓ − V −1 ( I + K)v, v, Γ
2
1
≥ v2V −1 − ( I + K)vV −1 vV −1
2
≥ (1 − cK ) v2V −1
= (1 − cK ) V −1 v, vΓ
1
≥ (1 − cK ) v2H 1/2 (Γ )
cV2
cV1
≥ (1 − cK ) vH 1/2 (Γ ) τv H −1/2 (Γ )
cV2
7.2 Neumann Boundary Value Problem 175
Remark 7.2. To describe the solution of the Dirichlet boundary value problem
(7.6) we have described four different boundary integral equations, and we
have shown their unique solvability. Depending on the application and on the
discretization scheme to be used, each of the above formulations may have
their advantages or disadvantages. In this book, we will mainly consider the
approximate solution of the variational formulation (7.9).
where we have to find the yet unknown Dirichlet datum γ0int u ∈ H 1/2 (Γ ).
From the second boundary integral equation of the Calderon system (7.3) we
obtain
(Dγ0int u)(x) = (1 − σ(x))g(x) − (K g)(x) for x ∈ Γ (7.19)
which is a first kind Fredholm boundary integral equation. Due to (6.17) we
have that u0 ≡ 1 is an eigensolution of the hypersingular boundary integral
operator, i.e. (Du0 )(x) = 0. Hence we have ker D = span {u0 }, and to ensure
the solvability of the boundary integral equation (7.19) we need to assume,
by applying Theorem 3.6, the solvability condition
Note that (ker D)0 is the orthogonal space which is induced by ker D, see
(3.15). From
2 2
1 , α} vH 1/2 (Γ ) ≥ ĉ1 vH 1/2 (Γ )
= min{c̄D D
∗
7.2 Neumann Boundary Value Problem 177
The solution of the boundary integral equation (7.28) is given by the Neumann
series
∞
γ0int u(x) = ([1 − σ]I − K) (V g)(x) for x ∈ Γ. (7.29)
=0
The convergence of the Neumann series (7.29) follows from the contraction
1/2
property (6.52) of ([1 − σ]I − K) in H∗ (Γ ) when considering the equivalent
178 7 Boundary Integral Equations
When using the indirect single layer potential (7.4) we finally obtain the
boundary integral equation
to find the unknown density w ∈ H −1/2 (Γ ) which is given via the Neumann
series
∞
w(x) = ((1 − σ)I − K ) g(x) for x ∈ Γ. (7.34)
=0
The convergence of the Neumann series (7.34) follows from the contraction
−1/2
property (6.55) of ((1−σ)I −K ) in H∗ (Γ ) when considering the equivalent
Sobolev norm · V .
Remark 7.3. For the solution of the Neumann boundary value problem (7.16)
again we have given and analyzed four different formulations of boundary
integral equations. As for the Dirichlet boundary value problem each of them
may have their advantages and disadvantages. Here we will mainly consider
the approximate solution of the modified variational formulation (7.25).
7.3 Mixed Boundary Conditions 179
Hence we have to find the yet unknown Dirichlet datum γ0int u(x) for x ∈ ΓN
and the Neumann datum γ1int u(x) for x ∈ ΓD . Keeping in mind the differ-
ent boundary integral formulations for both the Dirichlet and the Neumann
problems, there seems to be a wide variety of different boundary integral for-
mulations to solve the mixed boundary value problem (7.35). Here we will
only consider two formulations which are based on the representation formula
(7.36) of the direct approach.
The symmetric formulation [134] is based on the use of the first boundary
integral equation in (7.3) for x ∈ ΓD while the second boundary integral
equation in (7.3) is considered for x ∈ ΓN ,
Let gD ∈ H 1/2 (Γ ) and gN ∈ H −1/2 (Γ ) be suitable extensions of the given
boundary data gD ∈ H 1/2 (ΓD ) and gN ∈ H −1/2 (ΓN ) satisfying
gD (x) = gD (x) for x ∈ ΓD , gN (x) = gN (x) for x ∈ ΓN .
Inserting these extensions into the system (7.37) this gives the symmetric
formulation to find
1/2 (ΓN ),
:= γ0int u − gD ∈ H
u −1/2 (ΓD )
t := γ1int u − gN ∈ H
such that
(V
t)(x) − (K u gD (x) − (V gN )(x)
)(x) = (σI + K) for x ∈ ΓD ,
u)(x) + (K
(D t)(x) = ((1 − σ)I − K )
gN (x) − (D
gD )(x) for x ∈ ΓN .
(7.38)
180 7 Boundary Integral Equations
a( ; τ, v) = V
t, u , τ ΓD + K
t, τ ΓD − K u t, vΓN + D
u, vΓN ,
1 1
gD − V gN , τ ΓD + ( I − K )
F (τ, v) = ( I + K) gN − D
gD , vΓN .
2 2
Lemma 7.4. The bilinear form a(·; ·) of the symmetric boundary integral for-
−1/2 (ΓD ) × H
mulation is bounded and H 1/2 (ΓN )–elliptic, i.e.
a(t, u; τ, v) ≤ cA
2 (t, u)H -1/2 (ΓN ) (τ, v)H
-−1/2 (ΓD )×H -−1/2 (ΓD )×H
-1/2 (ΓN )
and
2
1 } (τ, v)H
a(τ, v; τ, v) ≥ min{cV1 , ĉD -−1/2 (Γ -1/2 (ΓN )
D )×H
−1/2 (ΓD ) × H
for all (t, u), (τ, v) ∈ H 1/2 (ΓN ) where the norm is defined by
(τ, v)2H
-−1/2 (Γ -1/2 (ΓN ) := τ 2H
-−1/2 (Γ + v2H
-1/2 (Γ .
D )×H D) N)
Proof. By using
we conclude the ellipticity of the bilinear form a(·, ·; ·, ·) from the ellipticity
estimates of the boundary integral operators V and D, see Theorem 6.22 for
d = 3 and Theorem 6.23 for d = 2, as well as (6.39). The boundedness of
the bilinear form a(·, ·; ·, ·) is a direct consequence of the boundedness of all
boundary integral operators.
Since the linear form F (τ, v) is bounded for (τ, v) ∈ H −1/2 (ΓD ) ×
H (ΓN ), the unique solvability of the variational formulation (7.39) follows
1/2
Let gD ∈ H 1/2 (Γ ) be some arbitrary but fixed extension of the given Dirichlet
datum gD ∈ H 1/2 (ΓD ). Then we have to find u 1/2 (ΓN )
:= γ0int u − gD ∈ H
such that
7.5 Exterior Boundary Value Problems 181
S
u, vΓN = gN − S
gD , vΓN (7.40)
1/2
is satisfied for all v ∈ H (ΓN ). Since the Steklov–Poincaré operator S :
1/2 (ΓN )–elliptic (see (6.48)) we con-
H 1/2 (Γ ) → H −1/2 (Γ ) is bounded and H
clude the unique solvability of the variational problem (7.40) from the Lax–
Milgram lemma (Theorem 3.4). If the Dirichlet datum γ0int u ∈ H 1/2 (Γ ) is
known, we can compute the complete Neumann datum γ1int u ∈ H −1/2 (Γ ) by
solving a Dirichlet boundary value problem.
is satisfied for all v ∈ H 1/2 (Γ ). By using (6.45) and the H 1/2 (Γ )–semi–
ellipticity of the hypersingular boundary integral operator D and assuming
κ(x) ≥ κ0 for x ∈ Γ we conclude
and therefore the H 1/2 (Γ )–ellipticity of the bilinear form a(·, ·). Again we
obtain the unique solvability of the variational problem (7.41) from the Lax–
Milgram lemma (Theorem 3.4).
Inserting the far field boundary condition (7.43) and taking the limit R → ∞
this results in the representation formula for x ∈ Ω c ,
u(x) = u0 − U ∗ (x, y)γ1ext u(y)dsy + γ1,y ext U ∗ (x, y)γ ext u(y)ds .
0 y
Γ Γ
To find the unknown Cauchy data again we can formulate different boundary
integral equations. The application of the exterior trace operator gives
Using the boundary integral equations of this system, the exterior Calderon
projection, we can formulate different boundary integral equations to handle
exterior boundary value problems with different boundary conditions. In par-
ticular for the exterior Dirichlet boundary value problem (7.42) and (7.43) we
can find the yet unknown Neumann datum γ1ext u ∈ H −1/2 (Γ ) as the unique
solution of the boundary integral equation
Note that the unique solvability of the boundary integral equation (7.44)
follows as for interior boundary value problems from the mapping properties
of the single layer potential V : H −1/2 (Γ ) → H 1/2 (Γ ).
7.6 Helmholtz Equation 183
then the single layer potential Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is not injective, i.e.
Moreover,
1
( I − Kk )γ1int uλ (x) = 0 for x ∈ Γ.
2
Proof. The assertion immediately follows from the direct boundary integral
equations
1
(Vk γiint uλ )(x) = ( I + Kk )γ0int uλ (x) = 0 for x ∈ Γ
2
and
1
( I − Kk )γiint uλ (x) = (Dk γ0int uλ )(x) = 0 .
2
Note that the exterior Dirichlet boundary value problem is uniquely solv-
able due to the Sommerfeld radiation condition. The solution is given by the
representation formula
u(x) = − Uk∗ (x, y)γ1ext u(y)dsy + γ1,yext U ∗ (x, y)g(y)ds
k y for x ∈ Ω c .
Γ Γ
then
1
( I + Kk )γ0int uµ (x) = 0 for x ∈ Γ.
2
Proof. The assertion immediately follows from the direct boundary integral
equation
1
( I + Kk )γ0int uµ (x) = (Vk γ1int uµ )(x) = 0 for x ∈ Γ.
2
7.7 Exercises
and
x 1
|x| , ∇u(x) − iku(x) = O |x|2 as |x| → ∞.
8
Approximation Methods
XM := span{ϕk }M
k=1 ⊂ X
M
uM := u k ϕk ∈ X M (8.2)
k=1
Note that we have used the same trial and test functions for the Galerkin–
Bubnov method.
It remains to investigate the unique solvability of the variational problem
(8.3), the stability of the approximate solutions uM ∈ XM as well as their
convergence for M → ∞ to the unique solution u ∈ X of the variational
problem (8.1). Due to XM ⊂ X we can choose v = vM ∈ XM in the variational
formulation (8.1). Subtracting the Galerkin–Bubnov problem (8.3) from the
continuous variational formulation (8.1) this gives the Galerkin orthogonality
M
uk Aϕk , ϕ = f, ϕ for = 1, . . . , M.
k=1
With
AM [, k] := Aϕk , ϕ , f := f, ϕ
for k, = 1, . . . , M this is equivalent to the linear system of algebraic equations
AM u = f (8.5)
M
M
M
M
(AM u, v) = AM [, k]uk v = Aϕk , ϕ uk v
k=1 =1 k=1 =1
M
M
= A u k ϕk , v ϕ = AuM , vM .
k=1 =1
Proof. The unique solvability of the variational problem (8.3) was already
discussed before. For the approximate solution uM ∈ XM of (8.3) we conclude
from the X–ellipticity of A
2
1 uM X ≤ AuM , uM = f, uM ≤ f X uM X
cA
and therefore we obtain the stability estimate (8.6). By using the X–ellipticity
and the boundedness of the linear operator A, and by using the Galerkin
orthogonality (8.4) we get for any arbitrary vM ∈ XM
2
1 u − uM X ≤ A(u − uM ), u − uM
cA
= A(u − uM ), u − vM + A(u − uM ), vM − uM
= A(u − uM ), u − vM
≤ cA
2 u − uM X u − vM X
BgX ≤ cB
2 gY for all g ∈ Y.
The generation of the linear system (8.5) then requires the computation of
N
gN = gi ψi ∈ YN = span{ψi }N
i=1 ⊂ Y.
i=1
M ∈ XM of the perturbed
Then we have to find an approximate solution u
variational problem
A
uM , vM = BgN , vM for all vM ∈ XM . (8.11)
= BN g
AM u (8.12)
Proof. When subtracting the perturbed variational problem (8.11) from the
Galerkin variational problem (8.10) this gives
A(uM − u
M ), vM = B(g − gN ), vM for all vM ∈ XM .
cA M 2X ≤ A(uM − u
1 uM − u M ), uM − u
M
= B(g − gN ), uM − u
M
≤ B(g − gN )X uM − u
M X
≤ cB
2 g − gN Y uM − u
M X .
cB
2
uM − u
M X ≤ g − gN Y .
cA
1
u − u
M X ≤ u − uM X + uM − u
M X
Subtracting the perturbed variational problem (8.13) from the Galerkin vari-
ational problem (8.3) we find
M −u
M 2X ≤ A(u
1 uM − u
cA M ), uM − u
M
− A)uM , uM − u
= (A M
− A)uM X uM − u
≤ (A M X
and therefore
1 M X ∗ .
uM − u
M X ≤ (A − A)u
cA
1
u − u
M X ≤ u − uM X + uM − u
M X
1 M X
≤ u − uM X + (A − A)u
cA
1
1 X + 1 [cA
≤ u − uM X + (A − A)u 2 +2 ] u − uM X
cA
A
c1
cA
1
and the assertion finally follows from Theorem 8.1 (Cea’s Lemma).
Theorem 8.4. Let u ∈ (ker B)⊥ be the unique solution of the operator equa-
tion Bu = g, and let uM ∈ XM be the unique solution of the variational prob-
lem (8.19). We further assume the discrete stability condition (8.20). Then
there holds the error estimate
cB
2
u − uM X ≤ 1 + inf u − vM X .
cB vM ∈XM
cB
2
PM vX ≤ vX .
cS
In particular, for the unique solution uM ∈ XM of the variational problem
(8.19) we obtain uM = PM u. On the other hand we have vM = PM vM for all
vM ∈ XM . Hence we have for an arbitrary vM ∈ XM
u − uM X = u − vM + vM − uM X = u − vM − PM (u − vM )X
cB
≤ u − vM X + PM (u − vM )X ≤ 1 + 2 u − vM X
cS
and therefore the assertion.
The convergence of the approximate solution uM → u ∈ X as M → ∞
then follows as for a Galerkin–Bubnov method from an approximation prop-
erty of the trial space XM .
It remains to establish the discrete stability condition (8.20). A possible
criterion is the following result due to Fortin [58].
and
RM qΠ ≤ cR qΠ for all q ∈ Π,
then there holds the discrete stability condition (8.20) with
cS = cS /cR .
8.5 Mixed Formulations 195
BvM , q BvM , RM q
cS vM X ≤ sup = sup
0=q∈Π qΠ 0=q∈Π qΠ
BvM , RM q BvM , qM
≤ cR sup ≤ cR sup ,
0=q∈Π RM qΠ 0=qM ∈ΠM qM Π
XM = span{ϕk }M
k=1 ⊂ X, ΠN = span{ψi }N
i=1 ⊂ Π.
Then the Galerkin variational formulation of the saddle point problem (8.21)
is to find (uM , pN ) ∈ XM × ΠN such that
f = f, ϕ , gj = g, ψj
We first consider the unique solvability of the linear system (8.23) from an
algebraic point of view. The dimension of the system matrix K in (8.23) is
N + M . With
rang AM ≤ M, rang BN ≤ min{M, N }
we find
rang K ≤ M + min{M, N } .
In particular for M < N we obtain
rang K ≤ 2M < M + N = dim K.
i.e. the linear system (8.23) is in general not solvable. Hence we have to define
the trial spaces XM and ΠN with care. The necessary condition M ≥ N
shows, that the trial space XM has to be rich enough compared with the trial
space ΠN .
To investigate the unique solvability of the Galerkin variational problem
(8.22) we will make use of Theorem 3.13. When considering the conform-
ing trial space XM ⊂ X we obtain from the X–ellipticity of A the positive
definiteness of the matrix AM , i.e.
2
(AM v, v) = AvM , vM ≥ cA
1 vM X > 0
BvM , qN AūM , vM
cS qN Π ≤ sup = sup ≤ cA
2 ūM X
0=v M ∈RM vM X 0=v M ∈RM vM X
and with
2 2
cA
2 1 cA
2
qN 2Π ≤ ūM 2X ≤ A (BN A−1
M BN q, q)
cS c1 cS
1 cB2 1
pN Π ≤ SN cA
f X + SN gΠ (8.26)
c1 1 c1
and
cB cB 1 cB
uM X ≤ 1 + S2N 2A f X + SN 2A gΠ . (8.27)
c1 c1 c1 c1
and therefore
1 0 B 1
pN Π ≤ c2 ūM X + gΠ .
cS1 N
198 8 Approximation Methods
Hereby, ū = A−1
M f ∈R
M
↔ ūM ∈ XM is the unique solution of the variational
problem
AūM , vM = f, vM for all vM ∈ XM .
From the X–ellipticity of A we then find
1
ūM X ≤ f X .
cA
1
Moreover,
2
1 uM X ≤ AuM , uM
cA
0 1
2 pN Π uM X
= f, uM − BuM , pN ≤ f X + cB
and therefore
1 0 1
uM X ≤ A
f X + cB
2 pN Π .
c1
With (8.26) we then obtain (8.27) .
Using the stability estimates (8.26) and (8.27) we also obtain an error
estimate for the approximate solution (uM , pN ) ∈ XM × ΠN .
Theorem 8.8. Let all assumptions of Theorem 8.7 be valid. For the unique
approximate solution (uM , pN ) ∈ XM × ΠN of the saddle point problem (8.22)
there holds the error estimate
u − uM X + p − pN Π ≤ c inf u − vM X + inf p − qN Π .
vM ∈XM qN ∈ΠN
Proof. When taking the difference of the continuous saddle point formulation
(8.21) with the Galerkin variational problem (8.22) for the conforming trial
spaces XM × ΠN ⊂ X × Π we obtain the Galerkin orthogonalities
A(u − uM ), vM + BvM , p − pN = 0
B(u − uM ), qN =0
for all (vM , qN ) ∈ XM × ΠN . Using Theorem 8.7 we find the unique solution
(ūM − uM , p̄N − pM ) ∈ XM × ΠN , and we obtain the stability estimates
p − pN Π ≤ p − p̄N Π + p̄N + pN Π
≤ (1 + c1 cB
2 ) p − p̄N Π + (c1 c2 + c2 c2 ) u − ūM X
A B
and
PM vX ≤ cP vX for all v ∈ X,
then there holds the discrete stability condition (8.25) with
cS = cS /cP .
is satisfied for all vM ∈ XM . Note that the variational problem (8.28) for-
mally coincides with the Galerkin–Bubnov formulation (8.3). However, since
we now consider the more general case of a coercive operator instead of an el-
liptic operator, the numerical analysis to establish suitable stability and error
estimates is different.
AwM , vM
cS wM X ≤ sup (8.29)
vM ∈XM ,
vM
X >0 vM X
1
uM X ≤ f X (8.30)
cS
and the error estimate
9 cA :
2
u − uM X ≤ 1+ inf u − vM X . (8.31)
cS vM ∈XM
Aw̄M , vM
cS w̄M X ≤ sup = 0
vM ∈XM ,
vM
X >0 vM X
AuM , vM
cS uM X ≤ sup
vM ∈XM ,
vM
X >0 vM X
f, vM
= sup ≤ f X
vM ∈XM ,
vM
X >0 vM X
1 cA
GM wX = wM X ≤ AwX ≤ 2 wX .
cS cS
In particular, we have uM = GM u for the solution of the Galerkin variational
formulation (8.28). Since GM is a projection, GM vM = vM for all vM ∈ XM ,
we then find
u − uM X = u − vM + GM vM − uM X
9 cA :
2
≤ u − vM X + GM (u − vM )X ≤ 1+ u − vM X
cS
for all vM ∈ XM . From this, the error estimate (8.31) follows.
8.6 Coercive Operators 201
It remains to validate the discrete stability condition (8.29). Note, that for
an X–elliptic operator A we then obtain
AwM , wM AwM , vM
1 wM X ≤
cA ≤ sup
wM X vM ∈XM ,
vM
X >0 vM X
Applying Cea’s lemma (cf. Theorem 8.1) for the X–elliptic operator D we
also find the stability estimate
1 cC
2
v̄M X ≤ D
CwM X ≤ D wM X
c1 c1
and therefore
9 cC :
2
wM − v̄M X ≤ wM X + vM X ≤ 1+ D
wM X
c1
Hence, the approximation property (8.8) of the trial space gives the conver-
gence v̄M ∈ v̄ in X for M → ∞.
Considering as test function vM = wM − v̄M we obtain
202 8 Approximation Methods
8.7 Exercises
8.1 Let X be a Hilbert space and let a(·, ·) : X × X → R be a symmetric
and positive definite bilinear form. For the approximation of the minimization
problem
1
F (u) = min F (v), F (v) = a(v, v) − f, v
v∈X 2
we introduce a finite–dimensional trial space
XM = span{ϕk }M
k=1 ⊂ X.
*
N
Ω = TN = τ (9.1)
=1
I(k) := { ∈ N : xk ∈ τ } for k = 1, . . . , M.
Moreover,
J() := {k ∈ N : xk ∈ τ } for = 1, . . . , N
is the index set of all nodes xk with xk ∈ τ . Note that dim J() = d + 1 in
the case of the finite elements τ considered here. Finally,
K(j) := { ∈ N : kj ∈ τ } for j = 1, . . . , K
204 9 Finite Elements
r r
@ @
JJ@
@
@ J@
r r J @
@
@ Jr @
HH
r @r r
H@r
t t
HH HH
HH HH
Ht Ht
t
S
tP P
t S
PP PP S
PP PP S
PPt PPSt
d := sup |x − y|
x,y∈τ
is the diameter of the finite element τ , which coincides with the longest edge
of the element τ . Obviously, for d = 1 we have
∆ = h = d .
9.1 Reference Elements 205
d ≤ cF r for = 1, . . . , N
where the constant cF does not depend on TN . For the two–dimensional case
d = 2 we then have
πr 2 ≤ ∆ = h2 ≤ d2 ≤ c2F r 2 ,
h = hmax := max h
=1,...,N
while
hmin := min h
=1,...,N
holds for all neighboring finite elements τj and τ . Here, two finite elements τ
and τj are called neighboring, if the average τ ∩ τ j consists either of a node,
an edge, or a triangle.
In the one–dimensional case d = 1 each finite element τ can be described
via a local parametrization, in particular for x ∈ τ and 1 , 2 ∈ J() we have
1 1−ξ
1
1
∆ = dsx = |det J | dξ = |det J | dξ2 dξ1 = |det J |
2
τ τ 0 0
and therefore
|detJ | = 2 ∆ . (9.6)
If we consider a function v(x) for x ∈ τ we can write
∇ξ v (ξ) = J ∇x v(x)
and therefore
∇x v(x) = J − ∇ξ v (ξ).
As for the one–dimensional case d = 1 we can show the following norm equiv-
alence estimates:
Lemma 9.1. For d = 2 and m ∈ N0 there hold the norm equivalence inequal-
ities
1
(2∆ )1−m ∇m 2L2 (τ ) ≤ ∇m
ξ v
2
x vL2 (τ ) ≤ cm (2∆ )
1−m
∇m 2L2 (τ )
ξ v
cm
(9.7)
where 2 m
cF
cm = .
π
Proof. For m = 0 the assertion follows directly from
208 9 Finite Elements
v2L2 (τ ) = |v(x)|2 dx = v (ξ)|2 |detJ | dξ = 2 ∆
| v 2L2 (τ ) .
τ τ
= 2∆ λmax (J −1 J − ) ∇ξ v2L2 (τ )
as well as
∇x v2L2 (τ ) ≥ 2∆ λmin (J −1 J − ) ∇ξ v2L2 (τ ) .
It is therefore sufficient to estimate the eigenvalues of the matrix J J . With
a := |x 2 − x 1 |, b := |x 3 − x 1 |, α = <)(x 3 − x 1 , x 2 − x 1 )
4∆2 4∆2
λ2 = ≥ 2 2
λ1 a +b
and therefore we conclude
4∆2
≤ λmin (J J ) ≤ λmax (J J ) ≤ a2 + b2 .
a2 + b2
9.1 Reference Elements 209
Moreover,
2c2F
a2 + b2 ≤ 2 d2 ≤ 2 c2F r 2 ≤ ∆ .
π
Hence we have
2π 2c2
2 ∆ ≤ λmin (J J ) ≤ λmax (J J ) ≤ F ∆
cF π
and the eigenvalues of the inverse matrix J −1 J − can be estimated as
π c2
2 (2∆ )−1 ≤ λmin (J −1 J − ) ≤ λmax (J −1 J − ) ≤ F (2∆ )−1 .
cF π
Hence we conclude the norm equivalence inequalities for m = 1. For m > 1,
the assertion follows by recursive applications of the above estimates.
In the three–dimensional case d = 3 the reference element τ is given by
the tetrahedron
τ = ξ ∈ R3 : 0 ≤ ξ1 ≤ 1, 0 ≤ ξ2 ≤ 1 − ξ1 , 0 ≤ ξ3 ≤ 1 − ξ1 − ξ2 . (9.8)
For x ∈ τ we then have the local parametrization
3
x = x 1 + ξi (x i+1 − x 1 ) = x 1 + J ξ for ξ ∈ τ
i=1
1 1−ξ
1 1−ξ
1 −ξ2
1
= |det J | dξ3 dξ2 dξ1 = |det J | (9.9)
6
0 0 0
and therefore
|det J | = 6 ∆ . (9.10)
As for the two–dimensional case we can write a function v(x) for x ∈ τ as
v(x) = v(x 1 + J ξ) = v (ξ) for ξ ∈ τ.
Again, the application of the chain rule gives
∇ξ v (ξ) = J ∇x v(x), ∇x v(x) = J − ∇ξ v (ξ)
and in analogy to Lemma 9.1 we have:
210 9 Finite Elements
Lemma 9.2. For d = 3 and m ∈ N0 there hold the norm equivalence inequal-
ities
c1 ∆ h−2m
∇m 2L2 (τ ) ≤ ∇m
ξ v
2 −2m
x vL2 (τ ) ≤ c2 ∆ h ∇m 2L2 (τ )
ξ v
α := <)(x 2 − x 1 , x 3 − x 1 ),
β := <)(x 2 − x 1 , x 4 − x 1 ),
γ := <)(x 3 − x 1 , x 4 − x 1 ).
λ1 + λ2 + λ3 = a2 + b2 + c2
λmax (J J ) ≤ a2 + b2 + c2 .
λ1 λ2 λ3 = det(J J ) = |det J |2 = 36 ∆2
36∆2
≤ λmin (J J ) ≤ λmax (J J ) ≤ a2 + b2 + c2 .
[a2 + b2 + c2 ]2
Since the finite element τ is assumed to be shape regular, we can estimate
the length of all edges by
8
2 2 2 2 2 2 3 9 2 2
a + b + c ≤ 3d ≤ 3cF r ≤ 3 c h
16π 2 F
and hence we obtain
8 8
4 3 256π 4 2 3 9 2 2
h ≤ λmin (J J ) ≤ λmax (J J ) ≤ 3 c h .
c4F 81 16π 2 F
c1 ∆ h−2m
∇m 2L2 (τ ) ≤ ∇m
ξ v
2 −2m
x vL2 (τ ) ≤ c2 ∆ h ∇m 2L2 (τ )
ξ v
d+1
vh (ξ) = vk ψk1 (ξ) for ξ ∈ τ. (9.13)
k=1
for d = 2
and for d = 3
d+1
vh (x) = vh (x 1 + J ξ) = v k ψk1 (ξ) for x ∈ τ , ξ ∈ τ. (9.14)
k=1
∆
d+1
∆ 2
d+1
v 2k ≤ vh 2L2 (τ ) ≤ v k .
(d + 1)(d + 2) d+1
k=1 k=1
d+1
d+1
vh 2L2 (τ ) = vh , vh L2 (τ ) = vi vj ψi (ξ)ψj (ξ)|detJ |dξ = (G v , v )
i=1 j=1 τ
where
∆
G = (Id+1 + ed+1 e
d+1 )
(d + 1)(d + 2)
is the local mass matrix and ed+1 = 1 ∈ Rd+1 . From the eigenvalues of the
matrix Id+1 + ed+1 e
d+1 ,
λ1 = d + 2, λ2 = · · · = λd+1 = 1,
∇x vh L2 (τ ) ≤ cI h−1
vh L2 (τ ) (9.15)
d+1
v (ξ) = v k ψk1 (ξ)
k=1
we obtain for d = 1
∇ξ v = v 2 − v 1
and therefore
and thus
10 1
∇ξ v 2L2 (τ ) = (v 2 − v 1 )2 + (v 3 − v 1 )2 + (v 4 − v 1 )2
6
10 2 1
≤ 2v 2 + 2v 23 + 2v 24 + 6v 21 ≤ 2 vh 2L∞ (τ ) .
6
Altogether we therefore have
for d = 2 by
and for d = 3 by
Note that linear form functions are associated to degrees of freedom at the
nodes xk ∈ τ , while the quadratic form functions are associated to the edge
mid points x∗kj . If the function vh is quadratic on τ then we can write
1
2 (d+1)(d+2)
vh (x) = vh (x 1 + J ξ) = v k ψk2 (ξ) for x ∈ τ , ξ ∈ τ. (9.16)
k=1
1
2 1
L2 (τ ) = ∆
ϕB [ξ(1 − ξ)]2 dξ = h .
30
0
Moreover,
1 2
d 1
∇x ϕB 2
L2 (τ ) = h−1
[ξ(1 − ξ)] dξ = h−1 .
dξ 3
0
Hence we conclude for the one–dimensional bubble function the local inverse
inequality √
∇x ϕB L2 (τ ) = 10 h−1
ϕ L2 (τ ) .
B
ψB = ξ1 ξ2 (1 − ξ1 − ξ2 ) for ξ ∈ τ
∇x ϕB c h−1
L2 (τ ) ≤ ϕ L2 (τ ) .
B
(9.18)
ψ(ξ) = ξ1 ξ2 ξ3 (1 − ξ1 − ξ2 − ξ3 )
and therefore
1
v B L2 (τ ) = 6∆ [ψB (ξ)]2 dξ = ∆
415800
τ
as well as
c
∇x v B 2L2 (τ ) ≤ c ∆ h−2 2
∇ξ ψB L2 (τ ) = ∆ h−2
,
15120
in particular we conclude the local inverse inequality (9.18) also for d = 3.
ϕ1k (x)
AB
BA
BA
BA
ϕ1k (x) BA
@
@ B A
@ B A
@ B A
@ @ A
@
AA
r r @r r @
@ xk
xk
@
@
M
vh (x) = vk ϕ1k (x).
k=1
Lemma 9.7. For vh ∈ Sh1 (TN ) there hold the spectral equivalence inequalities
⎛ ⎞ ⎛ ⎞
1
M 1 M
⎝ ∆ ⎠ vk2 ≤ vh 2L2 (TN ) ≤ ⎝ ∆ ⎠ vk2 .
(d + 1)(d + 2) d+1
k=1 ∈I(k) k=1 ∈I(k)
the upper estimate. The lower estimate follows in the same way.
Lemma 9.8. For a piecewise linear function vh ∈ Sh1 (TN ) there holds the
inverse inequality
N
∇x vh 2L2 (TN ) ≤ cI h−2 2
vh L2 (τ ) .
=1
To prove some approximation properties of the trial space Sh1 (TN ) we will
use error estimates of certain interpolation and projection operators. For a
globally continuous function v ∈ C(TN ) we define the interpolation in the
space of piecewise linear functions,
M
Ih v(x) := v(xk )ϕk (x) ∈ Sh1 (TN ). (9.20)
k=1
Lemma 9.9. Let v|τ ∈ H 2 (τ )be given. Then there holds the local error esti-
mate
v − Ih vL2 (τ ) ≤ c h2 |v|H 2 (τ ) .
Proof. For the error of the piecewise linear interpolation we first have, by
using the norm equivalence inequalities of Theorem 9.3,
v − Ih vL2 (τ ) ≤ c ∆
v − Iτ v L2 (τ ) ,
v L∞ (τ ) ≤ c
v H 2 (τ ) .
Iτ : H 2 (τ ) → L2 (τ )
≤ (I − Iτ )uL2 (τ ) wL2 (τ ) ≤ c uH 2 (τ ) wL2 (τ ) .
|f (u)| ≤
c wL2 (τ ) |u|H 2 (τ ) .
v L2 (τ ) ≤
(I − Iτ ) c |
v |H 2 (τ )
v |H 2 (τ ) ≤ ĉ h2 |v|H 2 (τ )
v − Ih vL2 (τ ) ≤ c ∆ |
N
v − Ih v2L2 (TN ) ≤ c h4 |v|2H 2 (τ ) . (9.21)
=1
N
v − Ih v2H 1 (TN ) ≤ c h2 |v|2H 2 (τ ) . (9.22)
=1
Qh u, vh L2 (TN ) = u, vh L2 (TN ) for all vh ∈ Sh1 (TN ). (9.23)
we conclude
and therefore
220 9 Finite Elements
On the other hand, again by using the Galerkin orthogonality (9.25) we have
as well as
u − Qh uL2 (TN ) ≤ c h2 vH 2 (TN ) .
By interpolating this estimate with the error estimate (9.26) this yields the
error estimate
u − Qh uL2 (TN ) ≤ c h vH 1 (TN ) . (9.28)
By Q1h : H 1 (TN ) → Sh1 (TN ) we denote the H 1 projection which is defined as
the unique solution of the variational problem
as well as
N
u − Q1h u2H 1 (TN ) ≤ u − Ih u2H 1 (TN ) ≤ c h2 |u|2H 2 (τ ) . (9.32)
=1
Hence we obtain the approximate property of the trial space Sh1 (TN ) of piece-
wise linear and continuous functions.
inf
1 (T )
u − vh H σ (TN ) ≤ c hs−σ |u|H s (TN ) . (9.33)
vh ∈Sh N
9.3 Trial Spaces 221
Lemma 9.11. For s ∈ (0, 1] let w ∈ H01 (TN ) be the uniquely determined
solution of the variational problem
w, vH 1 (TN ) = u − Q1h u, vH 1−s (TN ) for all v ∈ H 1 (TN ). (9.34)
Remark 9.12. In Lemma 9.11, the best possible value of s ∈ (0, 1] depends on
the regularity of the decomposition TN . If, for example, TN is convex, then
we obtain s = 1 [66]. In the case of a corner domain, see, for example, [49].
Note that due to Sh1 (TN ) ⊂ H 1+s (TN ) for s ∈ (0, 12 ) the H 1 projection Q1h u is
well defined also for functions u ∈ H 1−s (TN ) and s ∈ (0, 12 ). As in the proof
of Lemma 9.11 we then can conclude the stability estimate
Q1h uH 1−s (TN ) ≤ c uH 1−s (TN ) for all u ∈ H 1−s (TN ). (9.36)
Using the error estimates of Lemma 9.11 for s = 1 we can show the stability
of the L2 projection in H 1 (TN ).
222 9 Finite Elements
Proof. Let Q1h : H 1 (TN ) → Sh1 (TN ) ⊂ H 1 (TN ) be the H 1 projection as defined
in (9.29). By using the triangle inequality, the stability estimate (9.30), the
global inverse inequality (9.19), and the projection property Qh vh = vh for
all vh ∈ Sh1 (TN ) we obtain
Now the stability estimate follows from the error estimate (9.35) for Q1h and
from the stability estimate (9.30).
is valid.
Based on the trial space of piecewise linear and globally continuous func-
tions we can introduce trial spaces of locally higher polynomial degrees.
In the one–dimensional case d = 1 we can define quadratic basis functions
locally by
ϕ2 (x) = 4ξ(1 − ξ) for x = x 1 + ξ h ∈ τ .
An arbitrary function vh ∈ Sh2 (TN ) then can be written as
M
N
vh (x) = vk ϕ1k (x) + vM + ϕ2 (x).
k=1 =1
Therefore we have dim Sh2 (TN ) = M + N . For both the two–dimensional case
d = 2 and the three–dimensional case d = 3 we have to ensure the continuity
of the quadratic basis functions ϕ2 . Since the quadratic form functions are
9.3 Trial Spaces 223
defined locally with respect to the edges of the reference element, the support
of a global quadratic basis function consists of those finite elements which
share the corresponding edge. Denote by K the number of all edges of the
decomposition (9.1), then we can write for d = 2, 3 the global representation
M
K
vh (x) = vk ϕ1k (x) + vM +j ϕ2j (x)
k=1 j=1
Here, 1 , . . . , d+1 denote, as before, the indices of the associated global nodes,
while d+2 , . . . , 12 (d+1)(d+2) are the indices of the associated global edges, see
also Fig. 9.3 for d = 2.
Lemma 9.15. For vh ∈ Sh2 (TN ) there hold the spectral equivalence inequalities
2 2
dimS
h (TN ) dimS
h (TN )
with ⎧ ;
⎨ ∆ for k = 1, . . . , M,
∈I(k)
dk :=
⎩
∆k−M for k = M + 1, . . . , M + N
in the one–dimensional case d = 1, and
⎧ ;
⎪
⎪ ∆ for k = 1, . . . , M,
⎨ ∈I(k)
dk := ;
⎪
⎪ ∆ for k = M + 1, . . . , M + K
⎩
∈K(k−M )
when d = 2, 3.
Proof. First we will use the local spectral equivalence inequalities (9.17). For
d = 1 we have1
N
N 3
vh 2L2 (TN ) = vh 2L2 (τ ) ∆ v 2i
=1 =1 i=1
⎛ ⎞
M
N
= ⎝ ∆ ⎠ vk2 + 2
∆ vM + .
k=1 ∈I(k) =1
1
The equivalence A B means that there are positive constants c1 and c2 such
that c1 A ≤ B ≤ c2 A.
224 9 Finite Elements
and
N
u − Ih u H 1 (TN ) ≤ c h4 |u|2H 3 (τ ) .
=1
As in the case of piecewise linear basis functions we can show a global ap-
proximation property.
Theorem 9.16. Let u ∈ H s (TN ) with s ∈ [σ, 3] and σ = 0, 1. Then there
holds
inf
2
u − vh H σ (TN ) ≤ c hs−σ |u|H s (TN ) .
vh ∈Sh (TN )
By ShB (TN ) = span{ϕB } =1 we denote the global trial space of local bubble
N
Hence we have
⎡ ⎤2
c2d ∆ ⎣
QB 2 B 2
h vL2 (τ ) = |v | ϕ L2 (τ ) = v(x)dx⎦
∆2 cB
d
τ
with ⎧
⎨ 30 for d = 1,
cB
d = 2520 for d = 2,
⎩
415800 for d = 3.
By using c2d /cB
d < 2 for d = 1, 2, 3 and by applying the Cauchy–Schwarz
inequality we therefore obtain
⎡ ⎤2
2 ⎣ ⎦ 2
QB
h vL2 (τ ) ≤ v(x)dx ≤ dx [v(x)]2 dx = 2 v2L2 (τ ) .
∆ ∆
τ τ τ
Taking the sum over all elements this gives the desired stability estimate.
as defined in (9.23) is bounded (see (9.24)), and there holds the global error
estimate (9.27),
u − Qh uL2 (TN ) ≤ c h2 |u|H 2 (TN ) .
As already stated in Remark 9.14, the L2 projection
which admits a local error estimate. This can be done by using quasi interpo-
lation operators [42].
For any node xk of the locally uniform decomposition TN we define
*
ω k := τ
∈I(k)
Qkh u, vh L2 (ωk ) = u, vh L2 (ωk ) for all vh ∈ Sh1 (ωk ),
By using the local projection operators we can define the quasi interpolation
operator or Clement operator
M
(Ph u)(x) = (Qkh u)(xk ) ϕ1k (x).
k=1
Theorem 9.18. [27, 42, 139] For u ∈ H 1 (TN ) there holds the local error
estimate
u − Ph uL2 (τ ) ≤ c ĥk |u|H 1 (ωk ) for all = 1, . . . , N, (9.40)
k∈J( )
Proof. Let τ be an arbitrary but fixed finite element and let k ∈ J() be an
arbitrary fixed index. For x ∈ τ we can write
(Ph )(x) = (Qkh u)(x) + [(Qkh u)(xk ) − (Qkh )(xk )]ϕ1k (x).
k∈J( ),k=k
Therefore,
from which the error estimate (9.40) follows. The stability estimate (9.41) can
be shown in the same way.
9.5 Exercises
9.1 For an admissible decomposition of a bounded domain Ω ⊂ R2 into
triangular finite elements τ and for piecewise linear continuous basis functions
ϕk the mass matrix Mh is defined by
228 9 Finite Elements
Mh [j, k] = ϕk (x)ϕj (x)dx, j, k = 1, . . . , M.
Ω
Find a diagonal matrix Dh and positive constants c1 and c2 such that the
spectral equivalence inequalities
Compute the local mass matrix M as well as the minimal and maximal
eigenvalues of M .
10
Boundary Elements
*
N
ΓN = τ (10.2)
=1
r r r r r r r r r
r r
r r
r r
r r r r r r
r r
r r
r r
r r r r r
⎛⎞ ⎛1⎞ ⎛ 1 − 2ξ ⎞
ξ
⎜4⎟ ⎜ ⎟
χ1 (ξ) = ⎝ 4 ⎠ , χ2 (ξ) = ⎝ ⎠ , χ3 (ξ) = ⎝ 4 ⎠ ,
ξ 1
0
4 4
⎛ 1 ⎞ ⎛ξ−1⎞ ⎛ ⎞
− 0
⎜ 4 ⎟ ⎜ 4 ⎟
χ4 (ξ) = ⎝ ⎠ , χ5 (ξ) = ⎝ ⎠ , χ6 (ξ) = ⎝ ξ − 1 ⎠ .
1 − 2ξ 1
− 4
4 4
For j = 1, 2, 5, 6 the parameter domain Q = (0, 1) is decomposed into 4 equal
sized elements q j while for j = 3, 4 we have 8 elements q j to be used.
By {xk }Mk=1 we denote the set of all nodes of the boundary decomposition
ΓN . The index set I(k) describes all boundary elements τ where xk is a node,
while J() is the index set of all nodes xk describing the boundary element
τ . In the three–dimensional case d = 3 the boundary decomposition (10.2) is
called admissible, if two neighboring boundary elements share either a node
or an edge, see also Fig. 9.2. Analogous to (9.2) we compute by
∆ := dsx
τ
h := max h
=1,...,N
10.1 Reference Elements 231
d := sup |x − y|
x,y∈τ
hmin := min h .
=1,...,N
h
≤ cL for = 1, . . . , N
hj
1 /
d = h = ∆ = dsx = [x1 (ξ)]2 + [x2 (ξ)]2 dξ = |x 2 − x 1 | .
τ 0
then reads
x(ξ) = x 1 + ξ1 (x 2 − x 1 ) + ξ2 (x 3 − x 1 ) for ξ ∈ τ.
where
3 2
∂
E= xi (ξ) = |x 2 − x 1 |2 ,
i=1
∂ξ 1
3 2
∂
G= xi (ξ) = |x 3 − x 1 |2 ,
i=1
∂ξ 2
3
∂ ∂
F = xi (ξ) xi (ξ) = (x 2 − x 1 , x 3 − x 1 ) .
i=1
∂ξ1 ∂ξ2
and
∆ = dsx = det|χj (ξ)| dξ.
τ qj
10.2 Trial Spaces 233
With respect to the boundary decomposition (10.2) we now define trial spaces
of local polynomials. In particular we will consider the trial space Sh0 (Γ ) of
piecewise constant functions and the trial space Sh1 (Γ ) of piecewise linear
continuous functions. By considering appropriate interpolation and projection
operators we will prove certain approximation properties of these trial spaces.
Let
Sh0 (Γ ) := span{ϕ0k }N
k=1
be the space of functions which are piecewise constant with respect to the
boundary decomposition (10.2). The basis functions ϕ0k are given by
1 for x ∈ τk ,
ϕ0k (x) =
0 elsewhere.
N
uk ϕ0k , ϕ0 L2 (Γ ) = u, ϕ0 L2 (Γ ) for = 1, . . . , N.
k=1
Due to
∆k for k = ,
ϕ0k , ϕ0 L2 (Γ ) = ϕ0k (x)ϕ0 (x)dsx =
0 for k =
Γ
we obtain
1
uk = u(x)dsx for k = 1, . . . , N.
∆k
τk
Theorem 10.2. Let u ∈ H s (Γ ) be given for some s ∈ [0, 1], and let
Qh u ∈ Sh0 (Γ ) be the L2 projection as defined by (10.4). Then there hold the
error estimates
N
u − Qh u2L2 (Γ ) ≤ c h2s 2
k |u|H s (τk ) (10.5)
k=1
and
u − Qh uL2 (Γ ) ≤ c hs |u|H s (Γ ) . (10.6)
234 10 Boundary Elements
we obtain
u − Qh u2L2 (Γ ) = u − Qh u, u − Qh uL2 (Γ )
= u − Qh u, uL2 (Γ ) ≤ u − Qh uL2 (Γ ) uL2 (Γ )
and therefore
u − Qh uL2 (Γ ) ≤ uL2 (Γ )
which is the error estimate for s = 0.
We now consider s ∈ (0, 1). For x ∈ τk we have Qh u(x) = uk and therefore
1
u(x) − Qh u(x) = [u(x) − u(y)]dsy for x ∈ τk .
∆k
τk
By using the shape regularity (10.3) and ∆k = hd−1 k we can replace the
diameter dk and the area ∆k by the local mesh size hk ,
[u(x) − u(y)]2
|u(x) − Qh u(x)|2 ≤ cd−1+2s h 2s
dsy .
B k
|x − y|d−1+2s
τk
and by taking the sum over all boundary elements we obtain the error estimate
for s ∈ (0, 1).
10.2 Trial Spaces 235
and therefore
1 d
|u(x) − Qh u(x)| ≤ u(χj (t))|dt detχj (η)| dη
∆k dt
j j
qk qk
1
= |detχj (η)| dη |∇ξ u(χj (ξ))|dξ
∆k
j j
qk qk
= |∇ξ u(χj (ξ))|dξ.
j
qk
1 1
≤ χ |detχj (ξ)|dξ |u|2H 1 (τk ) = χ ∆k |u|2H 1 (τk ) .
c1 c1
j
qk
and by taking the sum over all boundary elements τk we finally obtain the
error estimate for s = 1 and d = 2.
236 10 Boundary Elements
1
1 d
= u(χj (η + t(ξ − η)))dt |detχj (η)| dη
∆k dt
j 0
qk
1
1
= (ξ − η) · ∇η u(χj (η + t(ξ − η)))dt |detχj (η)| dη.
∆k
j
qk 0
1
2
≤c |∇η u(χj (η + t(ξ − η)))| dt dη dξ
j j 0
qk qk
2
≤ c ∆k |∇η u(χj (η)| dξ = c ∆k |u|2H 1 (τk )
j
qk
10.2 Trial Spaces 237
By taking the sum over all boundary elements we finally get the error estimate
for s = 1 and d = 3,
Corollary 10.3. Let u ∈ H s (Γ ) be given for some s ∈ [0, 1]. For σ ∈ [−1, 0)
then there hold the error estimates
N
u − Qh u2H σ (Γ ) ≤ c h−2σ h2s 2
k |u|H s (τk )
k=1
and
u − Qh uH σ (Γ ) ≤ c hs−σ |u|H s (Γ ) . (10.8)
|u − Qh u, vL2 (Γ ) |
u − Qh uH σ (Γ ) = sup
0=v∈H −σ(Γ ) vH −σ (Γ )
|u − Qh u, v − Qh vL2 (Γ ) |
= sup
0=v∈H −σ(Γ ) vH −σ (Γ )
v − Qh vL2 (Γ )
≤ u − Qh uL2 (Γ ) sup .
0=v∈H −σ(Γ ) vH −σ (Γ )
By using the error estimate (10.5) for u − Qh uL2 (Γ ) and the estimate (10.6)
for v − Qh vL2 (Γ ) the assertion follows.
Altogether we can formulate the approximation property of the trial space
Sh0 (Γ ) of piecewise constant functions.
Theorem 10.4. Let σ ∈ [−1, 0]. For u ∈ H s (Γ ) with some s ∈ [σ, 1] there
holds the approximation property of Sh0 (Γ )
inf
0 (Γ )
u − vh H σ (Γ ) ≤ c hs−σ |u|H s (Γ ) . (10.9)
vh ∈Sh
Proof. For σ ∈ [−1, 0] and s ∈ [0, 1] the approximation property is just the
statement of Theorem 10.2 and Corollary 10.3. It remains to prove the ap-
proximation property for σ ∈ [−1, 0) and s ∈ [σ, 0).
For a given u ∈ H σ (Γ ) let Qσh u ∈ Sh0 (Γ ) ⊂ H σ (Γ ) ⊂ L2 (Γ ) be the H σ (Γ )
projection which is defined as the unique solution of the variational problem
u − Qσh uH σ (Γ ) ≤ uH σ (Γ ) .
I − Qσh H σ (Γ )→H σ (Γ ) ≤ 1.
u − Qσh uH σ (Γ ) ≤ u − Qh uH σ (Γ ) ≤ c h−σ uL2 (Γ ) .
I − Qσh H s (Γ )→H σ (Γ )
s−0 s−σ
≤ I − Qσh H σ (Γ )→H σ (Γ ) σ−0 I − Qσh L2 (Γ )→H σ (Γ ) −σ
s−σ
≤ (c h−σ ) −σ = c(s, σ) hs−σ .
This gives the approximation property for σ ∈ [−1, 0) and s ∈ [σ, 0).
Now we consider the case where Γj ⊂ Γ is an open boundary part of
Γ = ∂Ω, and Sh0 (Γj ) is the associated trial space of piecewise constant basis
functions. As in (10.6) there holds the error estimate
-σ (Γj ) ≤ c h
u − Qh uH s−σ
|u|H s (Γj ) .
inf
0 (Γ )
u − vh H
-σ (Γj ) ≤ c h
s−τ
|u|H s (Γj ) (10.10)
vh ∈Sh j
∆
d
∆ 2
d
v 2k ≤ vh 2L2 (τ ) ≤ v k .
d(d + 1) d
k=1 k=1
where
∆
G = (Id + ed e
d)
d(d + 1)
is the local mass matrix and ed = 1 ∈ Rd . The eigenvalues of the matrix
Id + ed e
d are given by
λ1 = d + 1, λ2 = · · · = λd = 1
Proof. Since the maximum of |vh | and therefore the vh L∞ (τ ) norm is equal
to some nodal value |vh (xk∗ )| = |vk∗ | for some xk∗ , the assertion follows
immediately from Lemma 10.5.
Lemma 10.7. For a function vh which is linear on τ there holds the local
inverse inequality
|vh |H 1 (τ ) ≤ cI h−1
vh L2 (τ ) .
240 10 Boundary Elements
qj
N
N
|vh |2H 1 (Γ ) = |vh |2H 1 (τ ) ≤ c h−2 2
vh L2 (τ )
=1 =1
|vh |H 1 (Γ ) ≤ c h−1 vh L2 (Γ ) .
vh H 1 (Γ ) ≤ c h−1 vh L2 (Γ ) ,
Analogous to the error estimates (9.21) and (9.22) we can estimate the interpo-
lation error of the piecewise linear interpolation operator Ih : H 2 (Γ ) → Sh1 (Γ )
as follows.
Lemma 10.8. Let v ∈ H 2 (Γ ) be given. Assume that Γ = ∂Ω is sufficiently
smooth where Ω ⊂ Rd . Let Ih v be the piecewise linear interpolation satisfying
Ih v(xk ) = v(xk ) at all nodes xk of the admissible boundary decomposition
(10.2). Then there hold the error estimates
N
v − Ih v2L2 (Γ ) ≤ c h4 |v|2H 2 (τ ) ≤ c h4 |v|2H 2 (Γ )
=1
and
N
v − Ih v2H 1 (Γ ) ≤ c h2 |v|2H 2 (τ ) ≤ c h2 |v|2H 2 (Γ ) .
=1
The piecewise linear interpolation requires, as in the case of finite elements, the
global continuity of the function to be interpolated. The function v ∈ H s (Γ )
is continuous for s ∈ ( d−1
2 , 2] and the following error estimate holds,
u − Qh uL2 (Γ ) ≤ uL2 (Γ ) .
On the other hand, by using Lemma 10.8 we also have the error estimate
N
u − Qh uL2 (Γ ) ≤ u − Ih uL2 (Γ ) ≤ c h4 |u|2H 2 (Γ ) ≤ c h2 |u|2H 2 (Γ )
=1
inf
1 (Γ )
u − vh H σ (Γ ) ≤ c hs−σ |u|H s (Γ ) . (10.14)
vh ∈Sh
h : L2 (Γ ) → Sh (Γ )
For a given u ∈ L2 (Γ ) we define the projection operator QB B
QB
h vH 1/2 (Γ )
≤ wh H −1/2 (Γ ) sup .
0=v∈L2 (Γ ) vL2 (Γ )
is satisfied for all v ∈ H01 (Ω). By using Theorem 4.3 we can state the unique
solvability of the above variational formulation.
Let
-
Xh := Sh1 (Ω) ∩ H01 (Ω) = span{ϕ1i }M
i=1
be the conformal trial space of piecewise linear and globally continuous basis
functions ϕ1k which are zero on the boundary ∂Ω. Note that the trial space is
defined with respect to some admissible decomposition Ω = ∪N =1 τ of Ω into
finite elements τ . Then the Galerkin variational problem of (11.2) is to find
u0,h ∈ Xh such that
244 11 Finite Element Methods
∇u0,h (x)∇vh (x)dx = f (x)vh (x)dx − ∇ug (x)∇vh (x)dx (11.3)
Ω Ω Ω
is satisfied for all vh ∈ Xh . By applying Theorem 8.1 (Cea’s Lemma) there ex-
ists a unique solution u0,h of the Galerkin variational problem (11.3) satisfying
the error estimate (8.7),
cA
2
u0 − u0,h H 1 (Ω) ≤ inf u0 − vh H 1 (Ω) .
cA
1 vh ∈Xh
u0 − u0,h H 1 (Ω) ≤ c hs−1 |u|H s (Ω) for u ∈ H s (Ω), s ∈ [1, 2]. (11.4)
we get
u0 − u0,h 2L2 (Ω) = [u0 (x) − u0,h (x)][u0 (x) − u0,h (x)]dx
Ω
= ∇w(x)∇[u0 (x) − u0,h (x)]dx
Ω
= ∇[w(x) − Q1h w(x)]∇[u0 (x) − u0,h (x)]dx
Ω
where Q1h : H01 (Ω) → Xh ⊂ H01 (Ω) is the H01 (Ω) projection which is defined
similar to (9.29). For w ∈ H 2 (Ω) we obtain from the approximation property
(9.32) of the trial space Xh the error estimate
Altogether we have
u0 − u
0,h H 1 (Ω) ≤ c1 hs−1 |u0 |H s (Ω) + c2 h |ug |H 2 (Ω)
0 1
≤ c hs−1 |u0 |H s (Ω) + gH 3/2 (Γ ) .
Theorem 11.2. Let u ∈ H s (Ω) for some s ∈ [1, 2] be the unique solution of
the Dirichlet boundary value problem (11.1). Let ug ∈ H 2 (Ω) be an appropri-
ate chosen extension of the given Dirichlet datum g. Let uh be the approximate
solution of (11.1) which is defined by (11.8). Then there holds the error esti-
mate 0 1
u − uh H 1 (Ω) ≤ c hs−1 |u0 |H s (Ω) + ug H 2 (Ω) (11.9)
11.1 Dirichlet Boundary Value Problem 247
≤ c h wH 2 (Ω) ug − Ih ug H 1 (Ω)
≤ c h u0 − u
0,h L2 (Ω) ug − Ih ug H 1 (Ω) .
For the remaining second term we have, by applying integration by parts, for
w ∈ H01 (Ω) ∩ H 2 (Ω),
∇w(x)∇[Ih ug (x) − ug (x)]dx = − ∆w(x)[Ih ug (x) − ug (x)]dx
Ω Ω
≤ wH 2 (Ω) ug − Ih ug L2 (Ω)
≤ c u0 − u
0,h L2 (Ω) ug − Ih ug L2 (Ω) .
Altogether we have
u0 − u
0,h L2 (Ω)
0 1
0,h H 1 (Ω) + ug − Ih ug H 1 (Ω) + ug − Ih ug L2 (Ω) .
≤ c1 h u0 − u
For u0 , ug ∈ H 2 (Ω) the assertion now follows from the error estimates
for u0 − u0,h H 1 (Ω) as well as from the interpolation error estimates for
ug − Ih ug L2 (Ω) and ug − Ih ug H 1 (Ω) .
11.1 Dirichlet Boundary Value Problem 249
The Galerkin variational problem (11.7) to find the coefficients ūi for
- is equivalent to an algebraic system of linear equations Ah ū = f
i = 1, . . . , M
with the stiffness matrix Ah defined by
Ah [j, i] = ∇ϕ1i (x)∇ϕ1j (x)dx
Ω
-
for all v ∈ RM ↔ vh ∈ Xh ⊂ H01 (Ω) positive definite. In particular we have
the following result:
-
Lemma 11.4. For all v ∈ RM ↔ vh ∈ Xh ⊂ H01 (Ω) there hold the spectral
equivalence inequalities
N
N
= ∇vh 2L2 (τ ) ≤ cI h−2 2
vh L2 (τ )
=1 =1
⎛ ⎞
-
N
M
≤c h−2 vk2 = c ⎝ hd−2 ⎠ vk2
∆
=1 k∈J( ) k=1 ∈I(k)
250 11 Finite Element Methods
and therefore the upper estimate. On the other hand, by using the H01 (Ω)
ellipticity of the bilinear form a(·, ·) and by changing to the L2 (Ω) norm we
get
2 2
(Ah v, v) = a(vh , vh ) ≥ cA
1 vh H 1 (Ω) ≥ c1 vh L2 (Ω)
A
⎛ ⎞
-
N
N
M
= cA
1 vh 2L2 (τ ) ≥ c ∆ vk2 = c ⎝ hd ⎠ vk2
=1 =1 k∈J( ) k=1 ∈I(k)
@ @ @ @ @ @ @ @ @
@ @@ @@ @@ @@
@ @ @@ @@ @@ @
@
@ @ @ @
@ @ @ @@ @@ @
@ @ @ @ @ @
@ @ @ @ @@ @@ @
@
@ @ @
@ @ @ @ @ @@ @
@ @ @ @ @ @ @
@ @ @ @ @ @ @@ @
@
@ @
@ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @
@ @
@ @ @ @ @ @ @ @
Fig. 11.1. Initial mesh (L = 0) and refined mesh L = 3.
The minimal and maximal eigenvalues and the resulting spectral condition
numbers of the associated finite element stiffness matrices are given in Table
11.1. Note that when choosing d = 2 and h = O(N −2 ) the results of Lemma
11.4 are confirmed. Thus, when using a conjugate gradient scheme to solve
the linear equation systems Ah u = f with the symmetric and positive def-
inite system matrices Ah we need to have an appropriate preconditioner to
bound the number of necessary iteration steps to reach a prescribed accuracy
independent of the system size. Note that preconditioned iterative schemes
are considered later in Chapter 13.
11.1 Dirichlet Boundary Value Problem 251
Next we will discuss the computation of the load vector f and the realization
of a matrix by vector multiplication with the stiffness matrix Ah as needed in
- we have by
the application of an iterative solution scheme. For j = 1, . . . , M
localization and parametrization, see Chapter 9,
fj = f (x)ϕ1j (x)dx = f (x)ϕ1j (x)dx
Ω ∈I(j) τ
= |det J | f (x 1 + J ξ)ψ 1j (ξ)dξ
∈τj τ
where j is the local index of the global node xj with respect to the finite
element τ . Hence we can reduce the computation of the global load vector
f to the computation of local load vectors f . In particular, for each finite
element τ we need to compute
f ,ι = |det J | f (x 1 + J ξ)ψι1 (ξ)dξ for ι = 1, . . . , d + 1.
τ
v ι := v ι + v ,ι .
Since the solution (11.14) of the Dirichlet boundary value problem is infinitely
often differentiable, we can apply Theorem 11.2 and Lemma 11.3 for s = 2.
Hence we obtain one as order of convergence when measuring the error in the
energy norm |u − uh |H 1 (Ω) . Moreover, when applying the Aubin–Nitsche trick
we get two as order of convergence when measuring the error in the L2 norm
u − uh L2 (Ω) . By eoc we denote the estimated order of convergence which
can be computed from
be the conforming trial space of piecewise linear and globally continuous basis
functions ϕ1k with respect to an admissible finite element mesh Ω = ∪N =1 τ .
Note that the basis functions {ϕ1k }M
k=1 build a partition of unity, i.e.
M
ϕ1k (x) = 1 for all x ∈ Ω. (11.18)
k=1
254 11 Finite Element Methods
ĉA
1
u − uh H 1 (Ω) ≤ inf u − vh H 1 (Ω) .
cA2 vh ∈Xh
on Γ = ∂Ω. Let
be the conforming finite element space of piecewise linear and globally contin-
uous basis functions ϕ1i . The restriction of Xh (Ω) onto Γ = ∂Ω then defines
a boundary element space
256 11 Finite Element Methods
Xh (Γ ) := Sh1 (Γ ) = span{φ1k }M
i=1 ⊂ H
Γ 1/2
(Γ )
of piecewise linear and continuous basis functions φ1k . We denote by MΩ the
number of all interior nodes xi ∈ Ω, and we have M = MΩ + MΓ as well as
Xh (Ω) = span{ϕ1i }M 1 M
i=1 ∪ span{ϕi }i=MΩ +1 .
Ω
In particular,
φ1i = γ0int ϕ1MΩ +i for i = 1, . . . , MΓ .
Moreover, let
−1/2
ΠH := span{ψk }N
k=1 ⊂ H (Γ )
denote a suitable trial space to approximate the Lagrange multiplier λ.
The Galerkin discretization of the saddle point problem (11.20) is to find
(uh , λH ) ∈ Xh × ΠH such that
γ0 uh (x)dsx γ0 vh (x)dsx + ∇uh (x)∇vh (x)dx − γ0int vh (x)λH (x)dsx
int int
Γ Γ Ω Γ
= f, vh Ω + g(x)dsx γ0int vh (x)dsx (11.21)
Γ Γ
γ0int uh (x)µH (x)dsx + λH (x)dsx µH (x)dsx
Γ Γ Γ
= g, µH Γ − f (x)dx µH (x)dsx
Ω Γ
Obviously,
0 for i = 1, . . . , MΩ ,
Bh [, i] =
B̄h [, i − MΩ ] for i = MΩ + 1, . . . , M
with
B̄h [, i] = ψ (x)φ1i (x)dsx
Γ
for i = 1, . . . , MΓ and = 1, . . . , N .
The matrix a a + Ah is by construction symmetric and positive definite
and therefore invertible. In particular, the first equation in (11.22) can be
solved for u to obtain
0 1−1 0 1
u = a a + Ah f + Bh λ .
Inserting this into the second equation of (11.22) we end up with the Schur
complement system
2 0 1−1 3 0 1−1
Bh a a + Ah Bh + b b λ = g − Bh a a + Ah f. (11.23)
The unique solvability of the Schur complement system (11.23) and therefore
of the linear system (11.22), and hence of the discrete saddle point problem
(11.21), now follows from Lemma 8.6 where we have to ensure the discrete
stability condition (8.25), i.e.
µH , γ0int vh Γ
cS µH H −1/2 (Γ ) ≤
sup for all µH ∈ ΠH . (11.24)
0=vh ∈Xh (Ω) vh H 1 (Ω)
This stability condition is first considered for the boundary element trial
spaces ΠH and Xh (Γ ).
µH , wh Γ
c̄S µH H −1/2 (Γ ) ≤ sup for all µH ∈ ΠH . (11.25)
0=wh ∈Xh (Γ ) wh H 1/2 (Γ )
258 11 Finite Element Methods
1/2
and JµH H 1/2 (Γ ) = µH H −1/2 (Γ ) . Let Qh JµH ∈ Xh (Γ ) be the unique
solution of the variational problem
1/2
Qh JµH , wh H 1/2 (Γ ) = µH , wh Γ for all wh ∈ Xh (Γ ).
1/2 1
(I − Qh )JµH H 1/2 (Γ ) ≤ µH H −1/2 (Γ )
2
is satisfied. Then we have
Remark 11.6. To establish the stability condition (11.25) one may also use
Fortin’s criterion (Lemma 8.9). Then we have to prove the boundedness of
the projection operator Q h : H 1/2 (Γ ) → Xh (Γ ) ⊂ H 1/2 (Γ ) which is defined
via the variational formulation
h u, µH Γ = u, µH Γ
Q for all µH ∈ ΠH .
By using the inverse trace theorem (Theorem 2.22) the stability condition
(11.25) implies
µH , wh Γ
c̄S µH H −1/2 (Γ ) ≤ cIT sup for all µH ∈ ΠH .
0=wh ∈Xh (Γ ) Ewh H 1 (Ω)
Rh vH 1 (Ω) ≤ cR vH 1 (Ω)
µH , wh Γ
c̄S µH H −1/2 (Γ ) ≤ cIT cR sup for all µH ∈ ΠH ,
0=wh ∈Xh (Γ ) Rh Ewh H 1 (Ω)
t t t t t
@ @ @ @ @ @ @ @ @ @
@ @
@ @ @ @
@
@ @ @@ @ @ @@
t @ @ @ @t
@ @ @@ @ @
@ @ @ @ @ @
@ @ @
@ @ @ @ @ @ @@
t@ @ @ @ @ @ @t
@ @ @ @ @ @ @@
@ @ @ @ @ @ @
@ @ @ @ @ @ @@
@
t @ @ @ @ @ @ @t
@ @
@ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @
t @ @t @ @t @ @t @ @t
Table 11.3. Results for a Finite Element Method with Lagrange multipliers.
1 0
Proof. For λ ∈ Hpw (Γ ) we define QH λ ∈ SH (Γ ) to be the L2 projection as
defined in (10.4). By using the triangle inequality and the inverse inequality
we obtain
The error estimate now follows from Theorem 10.2, Corollary 10.3, and by
using the error estimate (11.26).
When choosing h = 12 H we conclude by applying Lemma 11.7 the error
estimate
1
λ − λH 2L2 (Γ ) ≤ c1 H |u|2H 2 (Ω) + c2 H 2 λ2Hpw
1 (Γ )
4
and therefore an asymptotic order of convergence which is 0.5 when measur-
ing the error in the L2 norm. However, the numerical results in Table 11.3
indicate a higher order of convergence which is equal to 1. This preasymptotic
behavior may be explained by different orders of magnitude in the constants
1 2 2
4 c1 uH 1 (Ω) and c2 tH 1 (Γ ) .
pw
11.4 Exercises
11.1 Consider the Dirichlet boundary value problem
Compute the finite element stiffness matrix when using piecewise linear basis
functions with respect to a uniform decomposition of the interval (0, 1).
11.2 Show that the eigenvectors of the finite element stiffness matrix as de-
rived in Exercise 11.1 are given by the nodal interpolation of the eigenfunctions
as obtained in Exercise 1.6. Compute the associated eigenvalues and discuss
the behavior of the resulting spectral condition number.
11.3 Derive a two–dimensional Gaussian quadrature formula which integrates
cubic polynomials over the reference triangle
exactly.
12
Boundary Element Methods
Lu(x) = 0 for x ∈ Ω
Since the Cauchy data [γ0int u(x), γ1int u(x)] are given by the boundary con-
ditions only partially for x ∈ Γ , we need to find the remaining data from
the solution of suitable boundary integral equations. This chapter describes
boundary element methods as numerical discretization schemes to solve these
boundary integral equations approximately.
Sh0 (Γ ) = span{ϕ0k }N
k=1 ⊂ H
−1/2
(Γ )
be the trial space of piecewise constant basis functions ϕ0k . By using
N
th (x) = tk ϕ0k (x) ∈ Sh0 (Γ ) (12.4)
k=1
1 1 cW
2
th H −1/2 (Γ ) ≤ ( I + K)gH 1/2 (Γ ) ≤ gH 1/2 (Γ ) ,
cV1 2 cV1
cV2
J
t − th H −1/2 (Γ ) ≤ inf t|Γj − τhj H
-−1/2 (Γj ) . (12.7)
cV1 j=1 τhj ∈Sh0 (Γj )
From the approximation property (10.10) we then obtain the a priori error
estimate
1
t − th H −1/2 (Γ ) ≤ c hs+ 2 |t|Hpw
s (Γ ) (12.8)
Remark 12.1. When considering a direct boundary integral approach for the
Dirichlet boundary value problem of the Laplace equation, the solution
t ∈ H −1/2 (Γ ) is the exterior normal derivative t(x) = n(x) · ∇u(x) for x ∈ Γ .
When Γ = ∂Ω is the boundary of a domain with corners or edges, the ex-
terior normal vector n(x) is discontinuous and therefore the exterior normal
derivative t(x) is also discontinuous. Hence we have t ∈ H 2 −ε (Γ ) for any
d−1
arbitrary small ε > 0. When using globally continuous trial spaces, e.g. the
trial space Sh1 (Γ ) of piecewise linear basis functions, we can not apply the
error estimate (12.7) due to the global definition of Sh1 (Γ ), but we can apply
the global error estimate (12.6) for s < 12 (d − 1). Hence, when using higher
order basis functions they have to be discontinuous globally.
t − th L2 (Γ ) ≤ t − Qh tL2 (Γ ) + Qh t − th L2 (Γ )
≤ t − Qh tL2 (Γ ) + c h−1/2 Qh t − th H −1/2 (Γ )
0 1
≤ t − Qh tL2 (Γ ) + c h−1/2 t − Qh tH −1/2 (Γ ) + t − th H −1/2 (Γ ) .
The assertion now follows from the error estimates (10.6), (10.8), and (12.8).
After the approximate solution th ∈ Sh0 (Γ ) is determined as the unique
solution of the Galerkin variational problem (12.5), we obtain from (12.2) an
approximate solution of the Dirichlet boundary value problem (12.1), i.e. for
∈ Ω we have
x
∗ int U ∗ (
(
u x) = U (x, y)th (y)dsy − γ1,y x, y)g(y)dsy . (12.10)
Γ Γ
|u(
x) − u x)| ≤ U ∗ (
( x, ·)H −σ (Γ ) t − th H σ (Γ ) . (12.11)
t − th , vΓ
t − th H σ (Γ ) = sup .
0=v∈H −σ (Γ ) vH −σ (Γ )
we obtain
t − th , V wΓ
t − th H σ (Γ ) = sup
0=w∈H −1−σ (Γ ) V wH −σ (Γ )
V (t − th ), w − Qh wΓ
= sup .
0=w∈H −1−σ (Γ ) V wH −σ (Γ )
w − Qh wH −1/2 (Γ )
t − th H σ (Γ ) ≤
c t − th H −1/2 (Γ ) sup .
0=w∈H −1−σ (Γ ) wH −1−σ (Γ )
t − th H σ (Γ ) ≤ ĉ h−1/2−σ t − th H −1/2 (Γ ) .
t − th H −2 (Γ ) ≤ c h3 |t|Hpw
1 (Γ ) .
Moreover, when using (12.11) this gives the pointwise error estimate
|u(
x) − u x)| ≤ c h3 U ∗ (
( x, ·)H 2 (Γ ) |t|Hpw
1 (Γ ) .
Theorem 12.4. Let u ∈ H 1 (Ω) be the weak solution of the Dirichlet bound-
ary value problem (12.1), and let u ∈ H 1 (Ω) be the approximate solution as
defined via the representation formula (12.10). Then there holds the global
error estimate
u − uH 1 (Ω) ≤ c t − th H −1/2 (Γ ) . (12.13)
Hence we have
a(u0 − u
0 , v) = a(E(
g − g), v) for all v ∈ H01 (Ω).
cA 0 2H 1 (Ω) ≤ a(u0 − u
1 u0 − u 0 , u0 − u
0 ) = a(E(
g − g), u0 − u
0 )
2 E(
≤ cA g − g)H 1 (Ω) u0 − u
0 H 1 (Ω) ,
and therefore
cA
2
u0 − u
0 H 1 (Ω) ≤ E(
g − g)H 1 (Ω) .
cA
1
By applying the triangle inequality, and the inverse trace theorem, we conclude
u − u
H 1 (Ω) ≤ u0 − u
0 H 1 (Ω) + E(g − g)H 1 (Ω)
cA
≤ 1 + 2A E( g − g)H 1 (Ω)
c1
cA
≤ cIT 1 + 2A g − gH 1/2 (Γ )
c1
cA
≤ cIT 1 + 2A cV2 t − th H −1/2 (Γ )
c1
and therefore the claimed error estimate.
1
If t ∈ Hpw (Γ ), i.e. u ∈ H 5/2 (Ω), is satisfied, then from (12.13) we obtain
for the approximate solution u the error estimate in the energy space H 1 ,
Remark 12.5. When using lowest order, i.e. piecewise constant boundary el-
ements, and when assuming u ∈ H 5/2 (Ω) for the solution of the Dirichlet
boundary value problem, then the approximate solution u as defined via the
representation formula (12.10) converges with a convergence rate of 1.5. In
contrast, a lowest order, i.e. piecewise linear finite element solution converges
with a convergence rate of 1.0, see the error estimate (11.10), where we have
to assume u ∈ H 2 (Ω) only.
When inserting (12.4) into the Galerkin variational formulation (12.5), and
when choosing τh = ϕ0 as a test function, this gives
N
1
tk V ϕ0k , ϕ0 Γ = ( I + K)g, ϕ0 Γ for = 1, . . . , N.
2
k=1
By using
1
Vh [, k] = V ϕ0k , ϕ0 Γ ,
f = ( I + K)g, ϕ0 Γ
2
for k, = 1, . . . , N the variational problem is equivalent to a linear system of
algebraic equations,
Vh t = f (12.15)
where the stiffness matrix Vh is symmetric and positive definite, see Sec. 8.1.
12.1 Dirichlet Boundary Value Problem 269
Lemma 12.6. Let the boundary mesh be globally quasi–uniform. Then, for
all w ∈ RN ↔ wh ∈ Sh0 (Γ ) there hold the spectral equivalence inequalities
c1 hd w22 ≤ (Vh w, w) ≤ c2 hd−1 w22
by using ∆ = hd−1
.
To prove the lower estimate we consider an arbitrary but fixed w ∈ RN ↔
wh ∈ Sh0 (Γ ). The L2 projection QB
h wh ∈ Sh (Γ ) onto the space of local bubble
B
By using duality, the inverse inequality (10.16) in ShB (Γ ), and the stability
estimate (10.18) we further get
and therefore
N
(Vh w, w) ≥ c h wh 2L2 (Γ ) = c h w 2 ∆
=1
κ2 (Vh ) ≤ c h−1 .
elements τ while h = O(N −1 ) is the mesh size. In Table 12.1 there are given
the minimal and maximal eigenvalues of the stiffness matrix Vh as well as the
resulting spectral condition number κ2 (Vh ).
The computation of the load vector f requires the evaluation of
1
f = ( I + K)g(x)dsx for = 1, . . . , N
2
τ
where the double layer potential K is applied to the given Dirichlet data g. If
g is replaced by some piecewise linear approximation gh ∈ Sh1 (Γ ), we obtain
M
1
f = gi ( I + K)ϕ1i , ϕ0 Γ ,
i=1
2
in particular
1
f = ( Mh + Kh )g
2
where
Mh [, i] = ϕ1i , ϕ0 Γ , Kh [, i] = Kϕ1i , ϕ0 Γ
for i = 1, . . . , M and = 1, . . . , N . Hence we have to find the solution vector
t ∈ RN of the linear system
1
Vh
t = ( Mh + Kh )g.
2
1
≤ c1 hs+ 2 |t|Hpw
s (Γ ) + c2 g − gh 1/2
H (Γ ) (12.17)
M
gh (x) = Ih g(x) = g(xi )ϕ1i (x) ∈ Sh1 (Γ )
i=1
For the error estimate t − th H −1/2 (Γ ) in the energy space it is not impor-
tant whether the given data are replaced by some interpolation or by some
projection. However, this changes when considering error estimates in lower
Sobolev spaces, as they are used to obtain error estimates for the approximate
solution as defined via the representation formula (12.10).
272 12 Boundary Element Methods
t −
th H σ (Γ ) ≤ c1 hs−σ |t|Hpw
s (Γ ) + c2 h
−σ−1
|g|H (Γ )
V (t −
th ), wΓ
t −
th H σ (Γ ) = sup .
0=w∈H −1−σ (Γ ) V w H −σ (Γ )
V (t −
th ), w − Qh wΓ
sup ≤ c hs−σ |t|Hpw
s (Γ )
0=w∈H −1−σ (Γ ) V w H −σ (Γ )
and therefore
12.1 Dirichlet Boundary Value Problem 273
t −
th H σ (Γ ) ≤ c1 hs−σ |t|Hpw
s (Γ ) + c2 h
−σ−1
|g|H (Γ ) + c3 g − gh H 1+σ (Γ ) .
Table 12.2. Error and order of convergence, Dirichlet boundary value problem.
interpolation L2 projection
N |u(x̂) − û(x̂)| eoc |u(x̂) − û(x̂)| eoc
8 2.752 –2 6.818 –3
16 5.463 –3 2.33 5.233 –4 3.70
32 1.291 –3 2.08 6.197 –5 3.08
64 3.147 –4 2.04 6.753 –6 3.20
128 7.780 –5 2.02 7.587 –7 3.15
256 1.935 –5 2.01 8.878 –8 3.10
512 4.827 –6 2.00 1.070 –8 3.05
1024 1.205 –6 2.00 1.312 –9 3.03
2048 3.012 –7 2.00 1.620 –10 3.02
4096 7.528 –8 2.00 1.921 –11 3.08
theory: 2 3
Hence we have to find the yet unknown Dirichlet datum γ0int u ∈ H 1/2 (Γ ) as
the unique solution of the stabilized variational problem (7.24) such that
1
Dγ0int u, vΓ + γ0int u, weq Γ v, weq Γ = ( I − K )g, vΓ (12.24)
2
is satisfied for all v ∈ H 1/2 (Γ ). Let
Sh1 (Γ ) = span{ϕ1i }M
i=1 ⊂ H
1/2
(Γ )
be the trial space of piecewise linear and continuous basis functions ϕ1i . By
using
M
uh (x) = ui ϕ1i (x) ∈ Sh1 (Γ )
i=1
uh H 1/2 (Γ ) ≤ c gH −1/2 (Γ )
Due to
1
ker D = ker( I + K) = span{1} ⊂ Sh1 (Γ )
2
we find from the solvability assumption (1.17) the orthogonality relation
276 12 Boundary Element Methods
uh , weq Γ = 0
1/2
and therefore uh ∈ H∗ (Γ ).
By using the approximation property (10.14) we further obtain the error
estimate
γ0int u − uh H 1/2 (Γ ) ≤ c hs−1/2 |γ0int u|H s (Γ ) (12.26)
we obtain
Γ
γ0int u − uh , Dv
γ0int u − uh H σ (Γ ) = sup
0=v∈H 1−σ (Γ )
DvH −σ (Γ )
int u − uh ), v − Q vΓ
D(γ
1/2
0 h
= sup
0=v∈H 1−σ (Γ )
Dv H −σ (Γ )
1/2
where Qh : H 1/2 (Γ ) → Sh1 (Γ ) is the projection which is defined via the
variational formulation
1/2
Qh v, zh H 1/2 (Γ ) = v, zh H 1/2 (Γ ) for all zh ∈ Sh1 (Γ ).
Note that there holds the error estimate (10.13), i.e. for s ∈ [1, 2] we have
12.2 Neumann Boundary Value Problem 277
1/2
v − Qh vH 1/2 (Γ ) ≤ c hs−1/2 |v|H s (Γ ) .
≤ γ1int U ∗ (
x, ·)H 1 (Γ ) γ0int u − uh H −1 (Γ )
≤ c h3 γ1int U ∗ (
x, ·)H 1 (Γ ) |γ0int u|H 2 (Γ ) .
The Galerkin variational problem (12.25) is equivalent to a linear system of
algebraic equations,
(Dh + α a a )u = f (12.27)
where
1
Dh [j, i] = Dϕ1i , ϕ1j Γ , aj = ϕ1j , weq Γ ,
fj = ( I − K )g, ϕ1j Γ
2
h := Dh + α a a is
for i, j = 1, . . . , M . The stabilized stiffness matrix D
symmetric and positive definite, see also the general discussion in Sec. 8.1.
278 12 Boundary Element Methods
Lemma 12.9. Let the boundary mesh be globally quasi–uniform. Then, for
all v ∈ RM ↔ vh ∈ Sh1 (Γ ) there hold the spectral equivalence inequalities
h v, v) ≤ c2 hd−2 v22
c1 hd−1 v22 ≤ (D
g − Qh gH σ (Γ ) ≤ c hs−σ |g|Hpw
s (Γ )
1
(Dh + α a a )
u = ( Mh − Kh )g .
2
If R ⊂ Sh0 (Γ ) is satisfied, then we have
1/2
is satisfied, i.e. uh ∈ H∗ (Γ ).
The application of Theorem 8.2 (Strang Lemma) gives, when assuming
γ0int u ∈ H s (Γ ) and g ∈ Hpw
(Γ ) for some s ∈ [ 12 , 2] and ∈ [0, 1], respectively,
the error estimate
cD
2 cW
2
γ0int u − u
h H 1/2 (Γ ) ≤
inf1 γ0int u − vh H 1/2 (Γ ) +
g − gh H −1/2 (Γ )
cD v ∈Sh (Γ ) cD
1 h 1
1
≤ c1 hs−1/2 |γ0int u|H s (Γ ) + c2 h+ 2 |g|H (Γ ) . (12.29)
γ0int u − u
h H σ (Γ ) ≤ c1 hs−σ |γ0int u|H s (Γ ) + c2 h+1−σ |g|Hpw
(Γ )
where σ ∈ [0, 12 ].
int u − u
D(γ h ), vΓ
0
γ0int u − u
h H σ (Γ ) = sup .
0=v∈H 1−σ (Γ ) H −σ (Γ )
Dv
280 12 Boundary Element Methods
int u − u 1
D(γ0 h ), vh Γ = ( I − K )(g − gh ), vh Γ for all vh ∈ Sh1 (Γ ).
2
Therefore,
int u − u
D(γ
1/2
h ), v − Qh vΓ
0
γ0int u − u
h H σ (Γ ) ≤ sup
0=v∈H 1−σ (Γ )
DvH −σ (Γ )
1/2
( 12 I − K )(g − gh ), Qh vΓ
+ sup
0=v∈H 1−σ (Γ ) H −σ (Γ )
Dv
As in the proof of Theorem 12.8 we have
int u − u
D(γ h ), v − Qh vΓ
1/2
0
sup ≤ c1 hs−σ |γ0int u|H s (Γ )
0=v∈H 1−σ (Γ )
DvH −σ (Γ )
γ0int u − u
h H σ (Γ ) ≤ c1 hs−σ |γ0int u|H s (Γ ) + c2 h+1−σ |g|Hpw
(Γ )
+c3 g − gh H −1+σ (Γ ) .
1
Duh , vh Γ + ᾱ uh , 1Γ vh , 1Γ = ( I − K )g, vh Γ (12.30)
2
is satisfied for all vh ∈ Sh1 (Γ ). From the solvability assumption (1.17) we
1/2
then find uh ∈ H∗∗ (Γ ). All error estimates for the approximate solution
1
uh ∈ Sh (Γ ) follow as in Theorem 12.8. The variational problem (12.30) is
equivalent to a linear system of algebraic equations,
Dh + ᾱ ā ā u = f ,
where
āj = ϕ1j , 1Γ for j = 1, . . . , M.
Lu(x) = 0 for x ∈ Ω
γ0int u(x) = gD (x) for x ∈ ΓD ,
γ1int u(x) = gN (x) for x ∈ ΓN
a( ; τ, v) = F (τ, v)
t, u
−1/2 (ΓD ) × H
is satisfied for all (τ, v) ∈ H 1/2 (ΓN ). By using
282 12 Boundary Element Methods
ND
th (x) =
tk ϕ0k (x) ∈ Sh0 (ΓD ),
k=1
MN
h (x) =
u u 1/2 (ΓN )
i ϕ1i (x) ∈ Sh1 (ΓN ) ∩ H
i=1
a( h ; τh , vh ) = F (τh , vh )
th , u (12.31)
h 2H 1/2 (Γ ) +
u−u t−
th 2H −1/2 (Γ ) (12.32)
+
≤c inf u − vh 2H 1/2 (Γ ) +
inf
0 (Γ )
t − τh 2H −1/2 (Γ ) .
vh ∈Sh N
-1/2 (ΓN )
1 (Γ )∩H τh ∈Sh D
Assume that the solution u of the mixed boundary value problem (7.35) sat-
isfies u ∈ H s (Ω) for some s > 32 . Applying the trace theorem we obtain
γ0int u ∈ H s−1/2 (Γ ) as well as γ1int u ∈ Hpw
s−3/2
(Γ ). When assuming that the
extensions gD ∈ H s−1/2
(Γ ) and gN ∈ Hpw
s−3/2
(Γ ) are sufficiently regular, and
using the approximation properties of the trial spaces Sh0 (ΓD ) and Sh1 (ΓN ) we
obtain from (12.32) the error estimate
h 2H 1/2 (Γ ) +
u−u t− u|2H σ1 (Γ ) + c2 h2σ2 +1 |
th 2H −1/2 (Γ ) ≤ c1 h2σ1 −1 | t|2Hpw
σ2
(Γ )
where
1 1 1 3
≤ σ1 ≤ min 2, s − , − ≤ σ2 ≤ min 1, s − .
2 2 2 2
1
for all 2 ≤ σ ≤ min{2, s − 12 }. If u ∈ H 5/2 (Ω) is sufficiently smooth, we then
have
! "
h 2H 1/2 (Γ ) +
u−u t−
th 2H −1/2 (Γ ) ≤ c h3 u|2H 2 (Γ ) + |
| t|2Hpw
1 (Γ ) ,
i.e. the order of convergence for approximate solutions of the mixed bound-
ary value problem corresponds to those of the pure Dirichlet or Neumann
12.3 Mixed Boundary Conditions 283
Vh [, k] = V ϕ0k , ϕ0 ΓD , Dh [j, i] = Dϕ1i , ϕ1j ΓN , Kh [, i] = Kϕ1i , ϕ0 ΓD
S
u, vΓ = gN − S
gD , vΓN
1/2 (ΓN ). Recall that
is satisfied for all v ∈ H
1 1
S := D + ( I + K )V −1 ( I + K) : H 1/2 (Γ ) → H −1/2 (Γ )
2 2
is the Steklov–Poincaré operator. By using
MN
h (x) =
u u 1/2 (ΓN )
i ϕ1i (x) ∈ Sh1 (Γ ) ∩ H
i=1
= Dv + ( 1 I + K )wh
Sv (12.38)
2
we then define an approximation S of the Steklov–Poincaré operator S.
12.3 Mixed Boundary Conditions 285
Moreover, S is H
1/2 (ΓN )–elliptic,
vΓ ≥ cD v2 1/2
Sv, 1/2 (ΓN ),
for all v ∈ H
1 H (Γ )
from the H −1/2 (Γ ) ellipticity of the single layer potential V and from the
H 1/2 (ΓN ) ellipticity of the hypersingular boundary integral operator D.
286 12 Boundary Element Methods
= ( 1 I + K )(w − wh )
(S − S)v
2
we finally obtain
u − ûh H 1/2 (Γ )
+
≤c inf
u − vh H 1/2 (Γ ) + inf0 Su − τh H −1/2 (Γ ) .
-1/2 (ΓN )
1 (Γ )∩H
vh ∈Sh τh ∈Sh (Γ )
1 1
Sh = Dh + ( Mh + Kh )Vh−1 ( Mh + Kh ) (12.42)
2 2
12.4 Robin Boundary Conditions 287
ΓD = Γ ∩ {x ∈ R2 : x2 = 0},
and the remaining Neumann part ΓN = Γ \ΓD . The boundary data gD and
gN are given in such a way such that the solution of the mixed boundary value
problem (7.35) is
1
u(x) = − log |x − x∗ |, x∗ = (−0.1, −0.1) .
2
The boundary Γ = ∂Ω is decomposed uniformly into N boundary elements
τ of the same mesh width h. The boundary conditions on ΓD and ΓN are
interpolated by using piecewise linear and piecewise constant basis functions,
respectively. In Table 12.4 we give the errors and the related order of con-
vergence for the symmetric formulation (12.31) which confirm the theoretical
error estimates (12.33) and (12.34). The same holds true for the error esti-
mates (12.40) and (12.41) in the case of the symmetric approximation, for
which the numerical results are given in Table 12.5.
M
uh (x) = ui ϕ1i (x) ∈ Sh1 (Γ )
i=1
where
12.5 Exercises
12.1 Let τk be a one–dimensional boundary element of length hk with mid-
point x∗k . Compute the collocation matrix element
1
VhC [k, k] = − log |y − x∗k |dsy .
2π
τk
12.3 Using a simple midpoint rule the Galerkin matrix element of Exercise
12.2 may be approximated by
Vh [ + 1, k + 1] = Vh [, k] for k, = 1, . . . , N − 1
and
Vh [ + 1, 1] = Vh [, N ] for = 1, . . . , N − 1.
as well as of J , = 2, . . . , N .
12.8 Compute all eigenvalues and eigenvectors of a circulant matrix A. De-
scribe the inverse matrix A−1 .
13
Iterative Solution Methods
(AM pk , p ) = 0 for k, = 0, . . . , M − 1, k = .
292 13 Iterative Solution Methods
For an arbitrary given initial guess u0 ∈ RM we can write the unique solution
u ∈ RM of the linear system AM u = f as a linear combination of conjugate
vectors as
M−1
u = u0 − α p .
=0
Hence we have
M −1
AM u = AM u0 − α AM p = f ,
=0
and from the AM –orthogonality of the basis vectors p we can compute the
yet unknown coefficients from
(AM u0 − f , p )
α = for = 0, 1, . . . , M − 1.
(AM p , p )
k−1
uk := u0 − α p ∈ RM
=0
uk+1 := uk − αk pk for k = 0, 1, . . . , M − 1,
−1
and from the AM –orthogonality of the vectors {p }M
=0 we obtain
& '
k−1
0
AM u − α AM p − f , p
k
(AM u0 − f , pk ) =0 (AM uk − f , pk )
αk = = = .
(AM pk , pk ) (AM pk , pk ) (AM pk , pk )
If we denote by
rk := AM uk − f
the residual of the approximate solution uk we finally have
(rk , pk )
αk = . (13.1)
(AM pk , pk )
Initialize for k = 0:
p0 := w0
Compute for k = 0, 1, . . . , M − 2:
k
(AM wk+1 , p )
pk+1 := wk+1 − βk p , βk =
(Ap , p )
=0
span {p }k =0 = span {w }k =0 .
−1
It remains to define the initial vector system {w }M
=0 . One possibility is to
k k M
choose the unit basis vectors w := e = (δk+1, ) =1 [59]. Alternatively we
may find the basis vector wk+1 from the properties of the already constructed
vector systems {p }k =0 and {r }k =0 .
Lemma 13.1. For k = 0, 1, . . . , M − 2 we have
(rk+1 , p ) = 0 for = 0, 1, . . . , k.
(rk+1 , w ) = 0 for = 0, 1, . . . , k.
294 13 Iterative Solution Methods
−1
w = p + β −1,j pj .
j=0
Hence we obtain
−1
(rk+1 , w ) = (rk+1 , p ) + β −1,j (rk+1 , pj ),
j=0
w0 , w1 , . . . , wk , rk+1
are orthogonal to each other, and therefore linear independent. Since we need
only to know the search directions p0 , . . . , pk and therefore the initial vec-
tors w0 , . . . , wk to construct the approximate solution uk+1 and therefore the
residual rk+1 , we can define the new initial vector as
(rk+1 , r ) = 0 for = 0, . . . , k; k = 0, . . . , M − 2.
k−1
(rk , pk ) = (rk , rk ) + βk−1, (rk , p ) = (rk , rk ),
=0
(rk , rk )
αk = for k = 0, . . . , M − 1.
(AM pk , pk )
α > 0 for = 0, . . . , k.
(r +1 , r +1 ) = (r − α Ap , r +1 ) = (r , r +1 ) = 0
Now, by using wk+1 = rk+1 and by using the symmetry of the system matrix
AM = AM we can compute the nominator of the coefficient βk as
1 k+1
(AM wk+1 , p ) = (rk+1 , AM p ) = (r , r − r +1 )
α
⎧
⎪
⎨ 0 for < k,
= k+1
, rk+1 )
⎪
⎩−
(r
for = k.
αk
The recursion of the Gram–Schmidt orthogonalization algorithm now reduces
to
1 (rk+1 , rk+1 )
pk+1 = rk+1 − βkk pk where βkk = − .
αk (AM pk , pk )
On the other hand we have
and therefore
(rk+1 , rk+1 )
pk+1 = rk+1 + βk pk where βk = .
(rk , rk )
κ2 (Ah/2 ) − 1
FEM 2 κ2 (Ah ) − 1
FEM
.
1 κ2 (AFEM )+1 1
≈ ln . h
= ln qh .
2 κ2 (Ah ) − 1
FEM 2
13.1 The Method of Conjugate Gradients 297
Therefore, in the case of an uniform refinement step, i.e. halving the mesh
size h, the number of required iterations is doubled to reach the same relative
accuracy ε. As an example we choose ε = 10−10 . In Table 13.1 we give the
number of iterations of the conjugate gradient method to obtain the results
which were already presented in Table 11.2.
FEM BEM
L N κ2 (Ah ) Iter N κ2 (Vh ) Iter
2 64 12.66 13 16 24.14 8
3 256 51.55 38 32 47.86 18
4 1024 207.17 79 64 95.64 28
5 4096 829.69 157 128 191.01 39
6 16384 3319.76 309 256 381.32 52
7 65536 13280.04 607 512 760.73 69
8 262144 52592.92 1191 1024 1516.02 91
Theory: O(h−2 ) O(h−1 ) O(h−1 ) O(h−1/2 )
κ2 (ABEM
h ) = O(h−1 ) h/2 ) ≈ 2 κ2 (Ah
i.e. κ2 (ABEM BEM
).
satisfying
1/2 1/2 −1/2 1/2
CA = CA CA , CA := (CA )−1 .
Instead of the linear system AM u = f we now consider the equivalent system
298 13 Iterative Solution Methods
:= C −1/2 AM C −1/2
A A A
is again symmetric and positive definite. Hence we can apply the method
of conjugate gradients as described in Algorithm 13.2 to compute the trans-
1/2
= CA u. Inserting all the transformations we finally
formed solution vector u
obtain the preconditioned method of conjugate gradients, see Algorithm 13.3.
1/2
Note that for z = CA z we have
z 2A = (AC
1/2 1/2 2
A z, CA z) = (AM z, z) = zAM .
−1/2 k
Hence, for the approximate solution uk = CA we find the error estimate
u
2
q
uk − uAM ≤ u0 − uAM .
1 + q2k
13.2 A General Preconditioning Strategy 299
we
To bound the extremal eigenvalues of the transformed system matrix A
get from the Rayleigh quotient
z , z) z , z)
= min (A
λmin (A) ≤ max
(A
= λmax (A).
z , z)
z∈RM ( z , z)
z∈RM (
−1/2 −1/2 −1 cA
2
κ2 (CA ACA ) = κ2 (CA AM ) ≤ .
cA
1
−1
If the spectral condition number κ2 (CA AM ) of the preconditioned system
matrix can be bounded independent of the dimension M , i.e. independent of
the mesh size h, then there is a fixed number kε of required iterations to reach
a certain given relative accuracy ε.
and define
BM [, k] = Bψk , ψ , MM [, k] = ϕk , ψ for k, = 1, . . . , M.
Note that the Galerkin matrix BM is symmetric and positive definite, and
therefore invertible. Therefore we can define an approximation of the precon-
ditioning matrix CA by
A := M B −1 MM .
C (13.7)
M M
A is spec-
We need to prove that the approximated preconditioning matrix C
trally equivalent to CA , and therefore to AM .
Lemma 13.5. Let CA be the Galerkin matrix of B −1 as defined in (13.5),
A be the approximation as given in (13.7). Then there holds
and let C
A v, v) ≤ (CA v, v)
(C for v ∈ RM .
13.2 A General Preconditioning Strategy 301
Note that
(CA v, v) = B −1 vM , vM = w, vM = Bw, w. (13.8)
−1
In the same way we define w = BM MM v ↔ wM ∈ XM as the unique solution
of the Galerkin variational problem
BwM , zM = vM , zM for zM ∈ XM ,
Again,
A v, v) = (B −1 MM v, MM v) = (w, MM v) = wM , vM = BwM , wM .
(C M
(13.9)
Moreover we have the Galerkin orthogonality
B(w − wM ), zM = 0 for zM ∈ XM .
and therefore
BwM , wM ≤ Bw, w.
By using (13.8) and (13.9) this finally gives the assertion.
Note that Lemma 13.5 holds for any arbitrary conforming trial spaces
XM ⊂ X and XM ⊂ X . However, to prove the reverse estimate we need to
assume a certain stability condition of the trial space XM ⊂ X .
Then,
2
cB
1 A v, v)
cS (CA v, v) ≤ (C for all v ∈ RM .
cB
2
and hence
2 2
1 cB
2 1 cB
2
(CA v, v) ≤ wM 2X ≤ BwM , wM .
cB
1 cS cS cB
1
The bilinear form of the inverse single layer potential can then be written as
[v, weq Γ ]2
V −1 v, vΓ = V −1 v, vΓ + .
1, weq Γ
−1 vΓ ≤ 1 −1
1 V
cV1 cD v, vΓ ≤ Dv, V v, vΓ
4
for all v ∈ H 1/2 (Γ ).
By using Corollary 13.10 we now can define a preconditioner for the linear
system (12.15) of the Dirichlet boundary value problem, and for the sys-
tem (12.27) of the Neumann boundary value problem. The system matrix in
h := Dh + α a a where
(12.27) is D
for i, j = 1, . . . , M .
13.2 A General Preconditioning Strategy 305
Then,
≤ cQ wH −1/2 (Γ ) ,
h . In par-
defines a preconditioning matrix which is spectrally equivalent to D
ticular there hold the spectral equivalence inequalities
2
h v, v) ≤ 1 cV2
cV1 cD
1 (CD v, v) ≤ (D cQ V (CD v, v) for all v ∈ RM .
4 c1
In Table 13.2 the extremal eigenvalues and the resulting spectral condition
−1
numbers of the preconditioned system matrix CD Dh are listed for the L–
shaped domain as given in Fig. 10.1. For comparison we also give the corre-
sponding values in the case of a simple diagonal preconditioning which show
a linear dependency on the inverse mesh parameter h−1 .
By applying Corollary 13.10 we can use the Galerkin discretization of
the modified hypersingular boundary integral operator D as a preconditioner
for the discrete single layer potential Vh in (12.15). However, when using
piecewise constant basis functions to discretize the single layer potential, for
the Galerkin discretization of the hypersingular boundary integral operator
306 13 Iterative Solution Methods
CD
- = diag Dh CD −1
- = M̄h V̄h M̄h
−1 −1
L N λmin λmax κ(CD - Dh ) λmin λmax κ(CD - Dh )
0 28 9.05 –3 2.88 –2 3.18 1.02 –1 2.56 –1 2.50
1 112 4.07 –3 2.82 –2 6.94 9.24 –2 2.66 –1 2.88
2 448 1.98 –3 2.87 –2 14.47 8.96 –2 2.82 –1 3.14
3 1792 9.84 –3 2.90 –2 29.52 8.86 –2 2.89 –1 3.26
4 7168 4.91 –3 2.91 –2 59.35 8.80 –2 2.92 –1 3.31
5 28672 2.46 –4 2.92 –2 118.72 8.79 –2 2.92 –1 3.32
6 114688 1.23 –4 2.92 –2 237.66 8.78 –2 2.92 –1 3.33
Theory: O(h−1 ) O(1)
By construction we have
where Xh = Sh1 (Ω) ⊂ H 1 (Ω) is the trial space to be used for the Galerkin
−1 (Ω) which is induced by the
discretization of the operator A : H 1 (Ω) → H
bilinear form a(·, ·), i.e.
2 2
1 f H −1 (Ω) ≤ Bf, f Ω ≤ c2 f H −1 (Ω)
cB B
(13.13)
for all f ∈ H −1 (Ω) with some positive constants cB and cB . Such an op-
1 2
erator can be constructed when using an appropriately weighted multilevel
representation of L2 projection operators, see [28, 162].
For any trial space Vj ⊂ H 1 (Ω) let Qj : L2 (Ω) → Vj be the L2 projection
operator as defined in (9.23), i.e. Qj u ∈ Vj is the unique solution of the
variational problem
vj H 1 (Ω) ≤ cI h−1
j vj L2 (Ω) for all vj ∈ Vj . (13.15)
and therefore
Qk Qj v, vk L2 (Ω) = Qj v, vk L2 (Ω) = v, vk L2 (Ω) = Qk v, vk L2 (Ω)
From the inverse inequalities of the trial spaces Vk and by using the error
estimates of the L2 projection operators Qk we further obtain from Corollary
13.14:
Lemma 13.15. For all v ∈ H 1 (Ω) there hold the spectral equivalence inequal-
ities
∞
∞
c1 (Qk −Qk−1 )v2H 1 (Ω) ≤ B 1 v, vL2 (Ω) ≤ c2 (Qk −Qk−1 )v2H 1 (Ω) .
k=0 k=0
Proof. By using Lemma 13.12, 3., the triangle inequality, the error estimate
(13.14), and assumption (13.12) we have
∞
B 1 v, vL2 (Ω) = h−2 2
k (Qk − Qk−1 )vL2 (Ω)
k=0
∞
= h−2 2
k (Qk − Qk−1 )(Qk − Qk−1 )vL2 (Ω)
k=0
∞ !
≤ 2 h−2
k (Qk − I)(Qk − Qk−1 )v2L2 (Ω)
k=0
"
+ (I − Qk−1 )(Qk − Qk−1 )v2L2 (Ω)
∞
! "
≤ 2c h−2
k h2k (Qk − Qk−1 )v2H 1 (Ω) + h2k−1 (Qk − Qk−1 )v2H 1 (Ω)
k=0
∞
≤ c2 (Qk − Qk−1 )v2H 1 (Ω)
k=0
and therefore the upper estimate. To prove the lower estimate we get from
the global inverse inequality (13.15) for (Qk − Qk−1 )v ∈ Vk−1 , and by using
assumption (13.12),
310 13 Iterative Solution Methods
∞
∞
(Qk − Qk−1 )v2H 1 (Ω) ≤ c2I h−2 2
k−1 (Qk − Qk−1 )vL2 (Ω)
k=0 k=0
∞
≤c h−2 2
k (Qk − Qk−1 )vL2 (Ω)
k=0
1
= c B v, vL2 (Ω) .
The statement of Theorem 13.13 now follows from Lemma 13.15 and from the
following spectral equivalence inequalities.
Lemma 13.16. For all v ∈ H 1 (Ω) there hold the spectral equivalence inequal-
ities
∞
c̄1 v2H 1 (Ω) ≤ (Qk − Qk−1 )v2H 1 (Ω) ≤ c̄2 v2H 1 (Ω) .
k=0
To prove Lemma 13.16 we first need a tool to estimate some matrix norms.
Hence we have
13.2 A General Preconditioning Strategy 311
# $# $
|v |2 ≤ |A[, k]| 2α(k− ) |A[, k]| 2α( −k) u2k
∈I ∈I k∈I k∈I
# $ # $
≤ sup |A[, k]| 2
α(k− )
|A[, k]| 2
α( −k)
u2k
∈I
k∈I ∈I k∈I
# $ # $
= sup |A[, k]| 2
α(k− )
|A[, k]| 2
α( −k)
u2k
∈I
k∈I k∈I ∈I
# $ # $
≤ sup |A[, k]| 2α(k− ) sup |A[, k]| 2α( −k) u2k
∈I k∈I
k∈I ∈I k∈I
(Qi − Qi−1 )v, (Qj − Qj−1 )vH 1 (Ω) = (Qi − Qi−1 )v, Q1j (Qj − Qj−1 )vH 1 (Ω)
= Q1j (Qi − Qi−1 )v, (Qj − Qj−1 )vH 1 (Ω)
≤ Q1j (Qi − Qi−1 )vH 1 (Ω) (Qj − Qj−1 )vH 1 (Ω) .
Due to Vj = Sh1j (Ω) ⊂ H 1+σ (Ω) the H 1 projection as given in (9.29) is well
defined for u ∈ H 1−σ (Ω) and for σ ∈ (0, 12 ). Dependent on the regularity
312 13 Iterative Solution Methods
= cI h−s
j (Qi − Qi−1 )(Qi − Qi−1 )vH 1−s (Ω)
0
≤ cI h−s
j (Qi − I)(Qi − Qi−1 )vH 1−s (Ω)
1
+(I − Qi−1 )(Qi − Qi−1 )vH 1−s (Ω)
0 s 1
≤ c h−s
j hi + hsi−1 (Qi − Qi−1 )vH 1 (Ω)
≤
c 2s(j−i) (Qi − Qi−1 )vH 1 (Ω) .
With q := 2−s we obtain the strengthened Cauchy–Schwarz inequality.
Proof of Lemma 13.16: Let Q1j : H 1 (Ω) → Sh1j (Ω) ⊂ H 1 (Ω) be the H 1
projection as defined by the variational problem (9.29), in particular for a
given u ∈ H 1 (Ω) the projection Q1j u ∈ Vj is the unique solution of
For i < k we therefore have vi = (Q1i − Q1i−1 )v ∈ Vi−1 ⊂ Vk−1 , and thus
(Qk − Qk−1 )vi = 0. Hence, by interchanging the order of summation,
∞
∞
∞
(Qk − Qk−1 )v2H 1 (Ω) = (Qk − Qk−1 )vi , (Qk − Qk−1 )vj H 1 (Ω)
k=0 k=0 i,j=0
∞ min{i,j}
= (Qk − Qk−1 )vi , (Qk − Qk−1 )vj H 1 (Ω)
i,j=0 k=0
∞ min{i,j}
≤ (Qk − Qk−1 )vi H 1 (Ω) (Qk − Qk−1 )vj H 1 (Ω) .
i,j=0 k=0
13.2 A General Preconditioning Strategy 313
By using the global inverse inequality (13.15), the stability of the L2 projection
(see Remark 9.14), and applying some interpolation argument, we obtain from
assumption (13.12) for the already fixed parameter s ∈ (0, 1] the estimate
Moreover,
Hence we obtain
∞ ∞ min{i,j}
(Qk − Qk−1 )v2H 1 (Ω) ≤ c h−2s
k hsi hsj vi H 1 (Ω) vj H 1 (Ω) .
k=0 i,j=0 k=0
Finally,
∞
∞
vi 2H 1 (Ω) = (Q1i − Q1i−1 )v, (Q1i − Q1i−1 )vH 1 (Ω)
i=0 i=0
∞
= (Q1i − Q1i−1 )(Q1i − Q1i−1 )v, vH 1 (Ω)
i=0
∞
= (Q1i − Q1i−1 )v, vH 1 (Ω) = v, vH 1 (Ω) = v2H 1 (Ω) ,
i=0
Remark 13.19. For s ∈ [0, 32 ) we may define the more general multilevel oper-
ator
∞
B s := h−2s
k (Qk − Qk−1 )
k=0
Although the following considerations are only done for the special case s = 1,
these investigations can be extended to the more general case s ∈ [0, 32 ), too.
By using Theorem 13.13 the multilevel operator B 1 : H 1 (Ω) → H −1 (Ω)
1
is bounded and H (Ω)–elliptic. The inverse operator (B ) 1 −1
: H (Ω) →
−1
1 −1
H (Ω) is then bounded and H (Ω)–elliptic, in particular the spectral equiv-
alence inequalities (13.13) are valid. For the inverse operator (B 1 )−1 again a
multilevel representation can be given.
where s ∈ (− 12 , 12 ).
By using corollary 13.7 we can now establish the spectral equivalence of the
system matrix AhL with the discrete preconditioning matrix
A = M̄h B −1 M̄h
C L hL L
where
BhL [, k] = B −1 ϕL
k , ϕ L2 (Ω) ,
L
M̄hL [, k] = ϕL
k , ϕ L2 (Ω)
L
1
k , ϕ ∈ VL = ShL (Ω).
for all ϕL L
or,
u := Mh−1
L
r, w := BhL u, v := Mh−1
L
w.
By using the isomorphism u ∈ RML ↔ uhL ∈ VL we obtain for the components
of w = BhL u
ML
ML
w := BhL [, k]uk = B −1 ϕL
k , ϕ L2 (Ω) uk = B
L −1
L2 (Ω) .
uhL , ϕL
k=1 k=1
which is a finite sum due to Qk uhL = uhL for k ≥ L. For the components of
w = BhL u we then obtain
ML
−1
w = B L2 (Ω)
uhL , ϕL = zhL , ϕL
L2 (Ω) = zk ϕL
k , ϕ L2 (Ω) .
L
k=1
This is equivalent to
w = MhL z,
and therefore there is no need to invert the inverse mass matrix MhL when
computing the preconditioned residual,
v = Mh−1
L
w = Mh−1
L
MhL z = z .
L
zhL = h2k (Qk − Qk−1 )uhL
k=0
L−1
= h2L QL uhL + (h2k − h2k+1 )Qk uhL
k=0
L−1
= h2L ūhL + (h2k − h2k+1 )ūhk
k=0
where
Mk
ūhk = Qk uhL = ūk ϕk ∈ Vk
=1
we can define
L
z̄hL := h2k ūhk
k=0
Mk−1
Mk
z̄hk := z̄hk−1 + h2k ūhk = z̄ k−1 ϕk−1
+ h2k ūk ϕk .
=1 =1
Due to the inclusion Vk−1 ⊂ Vk we can write each basis function ϕk−1
∈ Vk−1
as a linear combination of basis functions ϕkj ∈ Vk ,
Mk
ϕk−1
= k
r ,j ϕkj for all = 1, . . . , Mk−1 .
j=1
Mk−1
Mk−1
Mk
Mk M k−1
z̄ k−1 ϕk−1
= z̄ k−1 k
r ,j ϕkj = z̄ k−1 r ,j
k
ϕkj .
=1 =1 j=1 j=1 =1
we obtain by induction
L
z̄ L = h2k Rk ūk .
k=0
@ @ @
@ @ @0
@ 0 @
1 @ 1
@ @ @
@ @1 @1
@ 2
@
2@
@ @ @
0 0 0 0 0
Fig. 13.1. Basis functions ϕk−1
and coefficients k
r,j (d = 2).
Mhk ūk = f k
where
Mhk [, j] = ϕkj , ϕk L2 (Ω) , f k = uhL , ϕk L2 (Ω) .
In particular for k = L we have
and therefore
ūL = Mh−1
L
r.
Due to
Mk
M
f k−1 = uhL , ϕk−1
L2 (Ω) = k
r ,j uhL , ϕkj L2 (Ω) = k
r ,j fjk
j=1 j=1
we get
f k−1 = Rk−1,k f k = Rk−1 r.
By recursion we therefore have
ūk = Mh−1
k
Rk r,
L
v = h2k Rk Mh−1
k
Rk r .
k=0
Taking into account the spectral equivalence of the mass matrices with the
diagonal matrices hdk I, see Lemma 9.7, we then obtain for the application of
the multilevel preconditioner
L
v = h2−d
k Rk Rk r. (13.19)
k=0
13.3 Solution Methods for Saddle Point Problems 319
CA = I CA = M̄hL Bh−1
L
M̄hL
−1 −1
L M λmin λmax κ(CA Ah ) λmin λmax κ(CA Ah )
1 13 2.88 –1 6.65 23.13 16.34 130.33 7.98
2 41 8.79 –2 7.54 85.71 16.69 160.04 9.59
3 145 2.42 –2 7.87 324.90 16.32 179.78 11.02
4 545 6.34 –3 7.96 1255.75 15.47 193.36 12.50
5 2113 1.62 –3 7.99 4925.47 15.48 202.94 13.11
6 8321 4.10 –4 8.00 19496.15 15.58 209.85 13.47
7 33025 ≈80000 15.76 214.87 13.63
8 131585 ≈320000 15.87 218.78 13.79
9 525313 ≈1280000 15.96 221.65 13.89
Theory: O(h−2 ) O(1)
In Table 13.3 we give the extremal eigenvalue and the resulting spec-
tral condition numbers of the preconditioned finite element stiffness matrix
−1
CA [a a +AhL ]. This preconditioner is also needed for an efficient solve of the
linear system (11.22), as it will be considered in the next section. The results
for the non–preconditioned system (CA = I) confirm the statement of Lemma
11.4, while the boundedness of the spectral condition of the preconditioned
systems coincides with the results of this section.
where the block A ∈ RM1 ×M1 is symmetric and positive definite, and
where D ∈ RM2 ×M2 is symmetric but positive semi–definite. Accordingly,
B ∈ RM1 ×M2 . Since the matrix A is assumed to be positive definite, we can
solve the first equation in (13.20) for u1 to obtain
320 13 Iterative Solution Methods
Inserting this into the second equation of (13.20) this results in the Schur
complement system
0 1
D + B A−1 B u2 = f 2 − B A−1 f 1 (13.21)
where
S = D + B A−1 B ∈ RM2 ×M2 . (13.22)
is the Schur complement. From the symmetry properties of the block matrices
A, B and D we conclude the symmetry of S, while, at this point, we assume
the positive definiteness of S.
We assume that for the symmetric and positive definite matrices A and
S = D + B A−1 B there are given some positive definite and symmetric
preconditioning matrices CA and CS satisfying the spectral equivalence in-
equalities
cA1 (CA x1 , x1 ) ≤ (Ax1 , x1 ) ≤ c2 (CA x1 , x1 )
A
(13.23)
for all x1 ∈ RM1 as well as
for all x2 ∈ RM2 . Hence, to solve the Schur complement system (13.21) we
can apply the CS preconditioned method of conjugate gradients (Algorithm
13.3). There, the matrix by vector multiplication sk = Spk for the Schur
complement (13.22) reads
Awk = Bpk .
This system can be solved either by a direct method, for example by the
Cholesky approach, or again by using a CA preconditioned method of con-
jugate gradients (Algorithm 13.3). Depending on the application under con-
sideration the Schur complement approach can be disadvantageous. Then, an
iterative solution strategy for the system (13.20) should be used. Possible iter-
ative solution methods for general non–symmetric linear systems of the form
(13.20) are the method of the generalized minimal residual (GMRES, [120]),
or the stabilized method of biorthogonal search directions (BiCGStab, [155]).
Here, following [26], we will describe a transformation of the block–skew
symmetric but positive definite system (13.20) leading to a symmetric and
positive definite system for which a preconditioned conjugate gradient ap-
proach can be used.
For the preconditioning matrix CA we need to assume that the spectral equiv-
alence inequalities (13.23) hold where
13.3 Solution Methods for Saddle Point Problems 321
cA
1 > 1 (13.25)
1 (CM x, x) ≤ (M x, x) ≤ c2 (CM x, x)
cM M
for all x ∈ RM1 +M2
where
322 13 Iterative Solution Methods
8
1 A 1 A
cM
1 = c2 [1 + c1 ] −
S
[c (1 + cS1 )]2 − cS1 cA
2 ,
2 4 2
8
1 1 A
cM
2 = cA [1 + cS2 ] + [c (1 + cS2 )]2 − cS2 cA
2 .
2 2 4 2
Proof. We need to estimate the extremal eigenvalues of the preconditioned
−1
system matrix CM M , in particular we have to consider the eigenvalue prob-
lem
& '& ' & '& '
−1 −1
ACA A − A (I − ACA )B x1 A − CA 0 x1
−1 −1 = λ .
B (I − CA A) D + B CA B x2 0 CS x2
Let λi be an eigenvalue with associated eigenvectors x1i and xi2 . From the first
equation,
−1 −1
(ACA A − A)xi1 + (I − ACA )Bxi2 = λi (A − CA )xi1 ,
For λi ∈ [1, cA
2 ] nothing is to be shown. Hence we only consider λi ∈ [1, c2 ]
A
Inserting this result into the second equation of the eigenvalue problem,
−1 −1
B (I − CA A)xi1 + [D + B CA B]xi2 = λi CS xi2 ,
this gives
−1 −1
−B CA (CA − A)(λi CA − A)−1 Bxi2 + [D + B CA B]xi2 = λi CS xi2 .
Due to
−1 −1
−CA (CA − A)(λi CA − A)−1 = −CA [λi CA − A + (1 − λi )CA ](λi CA − A)−1
−1
= (λi − 1)(λi CA − A)−1 − CA
this is equivalent to
λi − cA
2
(Ax1 , x1 ) ≤ ((λi CA − A)x1 , x1 )
cA
2
13.3 Solution Methods for Saddle Point Problems 323
cA
2
((λi CA − A)−1 x1 , x1 ) ≤ (A−1 x1 , x1 ) for all x1 ∈ Rn1 .
λi − cA
2
λi
(Sxi2 , xi2 ) ≤ λi (CS xi2 , xi2 )
cS2
= (Dxi2 , xi2 ) + (λi − 1)((λi CA − A)−1 Bxi2 , Bxi2 )
cA
2
≤ (Dxi2 , xi2 ) + (λi − 1) (A−1 Bxi2 , Bxi2 )
λi − cA
2
A λi − 1
≤ c2 i i
(Sx2 , x2 )
λi − cA 2
and therefore
λi λi − 1
≤ cA
2 ,
S
c2 λi − cA
2
i.e.
λ2i − cA
2 [1 + c2 ]λi + c2 c2 ≤ 0 .
S S A
2 − λi
cA
((A − λi CA )x1 , x1 ) ≤ (Ax1 , x1 )
cA
2
and therefore
cA
2
((A − λi CA )−1 x1 , x1 ) ≥ (A−1 x1 , x1 )
2 − λi
cA
for all x1 ∈ RM1 . Again, by using the spectral equivalence inequalities (13.24)
we conclude
324 13 Iterative Solution Methods
λi
(Sxi2 , xi2 ) ≥ λi (CS xi2 , xi2 )
cS1
= (Dxi2 , xi2 ) + (1 − λi )((A − λi CA )−1 Bxi2 , Bxi2 )
cA
≥ (Dxi2 , xi2 ) + (1 − λi ) A 2 (A−1 Bxi2 , Bxi2 )
c2 − λi
cA
2
≥ (1 − λi ) (Sxi2 , xi2 )
2 − λi
cA
and therefore
1 − λi λi
cA
2 ≤ S.
c2 − λi
A c1
This is equivalent to
λ2i − cA
2 [c1 + 1]λi + c1 c2 ≤ 0,
S S A
i.e.
λ− ≤ λi ≤ λ+
where 8
1 A 1 A
λ± = c2 [1 + c1 ] ±
S
[c (1 + cS1 )]2 − cS1 cA
2.
2 4 2
Summarizing we have
8
1 A 1 A
1 > λi ≥ c2 [1 + c1 ] −
S
[c (1 + cS1 )]2 − cS1 cA M
2 = c1 .
2 4 2
This completes the proof.
For the solution of the transformed linear system (13.26) Algorithm
13.3 of the preconditioned conjugate gradient approach can be applied. On
a first glance the multiplication with the inverse preconditioning matrix
−1 k+1
v k+1 = CM r , in particular the evaluation of v k+1
1 = (A−CA )−1 rk+1 seems
to be difficult. However, from the recursion of the residual,
rk+1 = rk − αk M pk , we find the representation
−1
rk+1
1 := rk1 − αk (ACA − I)(Apk1 − Bpk2 ).
L N M Schur CG BP CG
2 16 11 11 16
3 32 23 13 19
4 64 47 14 21
5 128 95 14 21
6 256 191 15 23
7 512 383 16 23
8 1024 767 16 23
9 2048 1535 16 24
When using the preconditioning matrix CD = M̄h V̄h−1 M̄h we then obtain the
estimate for the spectral condition number,
−1
κ2 (CD Dh ) ≤ c [1 + log |h|]2 .
where we assume diam Ω < 1 to ensure the invertibility of the single layer
potential V . When using an indirect single layer potential (7.4) the solution
of the above problem is given by
1
x) = −
u( log |
x − y|w(y)dsy for x ∈ Ω.
2π
Γ
The yet unknown density w ∈ H −1/2 (Γ ) is then given as the unique solution
of the boundary integral equation (7.12),
1
(V w)(x) = − log |x − y|w(y)dsy = g(x) for x ∈ Γ.
2π
Γ
2 2 2 2
Ω11 Ω12 Ω15 Ω16
Ω31 Ω41
Ω92 2
Ω10 2
Ω13 2
Ω14
Ω10
Ω32 Ω42 Ω72 Ω82
Ω11 Ω21
Ω12 Ω22 Ω52 Ω62
dλj = 2−λ , j = 1, . . . , 4λ .
The refinement strategy (14.4) is applied recursively until the edge length dL
j
of a box ΩjL on the finest level L is proportional to the mesh size h of the
globally quasi–uniform boundary mesh {τ }N =1 , i.e.
−L
dL
j = 2 ≤ cL h
induces a maximal number of boundary elements τ which are contained in
the box ΩjL . Then we find for the maximal level of the cluster tree
cL ln(1/h)
L ≥ .
ln 2
Since the boundary decomposition is assumed to be globally quasi–uniform,
this implies that the surface measure |Γ | is proportional to N h and therefore
we obtain
330 14 Fast Boundary Element Methods
L = O(ln N ) . (14.5)
To describe the clustering of the boundary elements {τ }N
we may consider
=1
the clustering of the associated element midpoints x̂ ∈ τ for = 1, . . . , N .
For j = 1, . . . , 4L we first collect all boundary elements τ where the midpoint
x̂ is in the box ΩjL in a cluster ωjL ,
*
ωLj := τ .
x̂ ∈ΩjL
The hierarchy (14.4) of boxes Ωjλ now transfers directly to a hierarchy of the
related clusters ωjλ , see Fig. 14.2,
4j
*
ω λ−1
j := ω λi for j = 1, . . . , 4(λ−1) , λ = L, . . . , 1. (14.6)
i=4(j−1)+1
ω10
P
@PPP
@ P
P
ω11 ω21 ω31 ω41
··
··
··
··
ω1L−2
P
@PPP
@ P
P
ω1L−1 ω2L−1 ω3L−1 ω4L−1
P
@PPP
@ P
P
ω1L ω2L ω3L ω4L
A
@
A@
τ1 · · · · · · τ1
denotes the number of boundary elements τ inside the cluster ωjλ . By con-
struction we have for each λ = 0, 1, 2, . . . , L
4
λ
Njλ = N . (14.7)
j=1
By
diam ωjλ := sup |x − y|
x,y∈ωjλ
we define the distance between the clusters ωiκ and ωjλ . A pair of clusters ωiκ
and ωjλ is called admissible, if
dist(ωiκ , ωjλ ) ≥ η max diam ωiκ , diam ωjλ (14.8)
L 4
λ
4 λ
4
L
4 L
λ,ij 1
Vh [, k] = − log |x − y|dsy dsx for τk ∈ ωiλ , τ ∈ ωjλ , (14.10)
2π
τ τk
see also Fig. 14.3. The sum is to be taken over all inadmissible clusters ωiL and
ωjL includes in particular the interaction of a cluster with itself and with all
neighboring clusters. Hence we denote this part as the near field of the stiffness
matrix Vh while the remainder, i.e. the sum over all maximally admissible
clusters is called the far field. A box ΩiL has maximal 8 direct neighbors
ΩjL , the associated cluster ωiL therefore has a certain number of inadmissible
332 14 Fast Boundary Element Methods
clusters ωjL where the number only depends on the parameter η > 1. All other
clusters are thus admissible and therefore they are included in the far field.
The near field part therefore contains only O(η 2 4L ) = O(η 2 N ) summands.
While the block matrices of the near field part can be evaluated directly as
in a standard boundary element method, the block matrices of the far field
part can be approximated by using low rank matrices which allow for a more
efficient application. The resulting matrices are called hierarchical matrices,
or H matrices. [72].
The basic idea for the derivation of fast boundary element methods is an
1
approximate splitting of the fundamental solution U ∗ (x, y) = − 2π log |x − y|
into functions which only depend on the integration point y ∈ ωi , and on the
λ
U∗ (x, y) = λ,j
fm λ,i
(x)gm (y) for (x, y) ∈ ωjλ × ωiλ . (14.11)
m=0
|U ∗ (x, y) − U∗ (x, y)| ≤ c(η, ) for (x, y) ∈ ωjλ × ωiλ , (14.12)
|Vhλ,ij [, k] − Vhλ,ij [, k]| ≤ c(η, ) ∆k ∆ for τk ∈ ωiλ , τ ∈ ωjλ . (14.14)
Hence, for all boundary elements τk ∈ ωiλ and τ ∈ ωjλ we need to compute
vectors defined by the entries
aλ,j
m, := f λ,j
m (x)ds x , bλ,i
m,k := λ,i
gm (y)dsy ,
τ τk
( + 1)(Niλ + Njλ ).
L 4
λ
4 λ
4
L
4 L
Due to (14.7) the total amount of work to store and to apply the approximate
stiffness matrix Vh is proportional to, by taking into account the near field
part,
( + 1)(η + 1)2 (L + 1)N + η 2 N. (14.16)
Instead of the original linear system Vh w = f we now have to solve the per-
turbed system Vh w
= f with an associated approximate solution w h ∈ Sh0 (Γ ).
The stability and error analysis of the perturbed problem is based on the
Strang lemma (Theorem 8.3). Hence we need to prove the positive definiteness
of the approximate stiffness matrix Vh , which will follow from an estimate of
the approximation error Vh − Vh .
Lemma 14.1. For a pair of maximally admissible clusters ωiλ and ωjλ the
error estimate (14.12) is assumed. Let Vh be the approximate stiffness matrix
as defined in (14.15). Then there holds the error estimate
L 4
4 λ λ
|((Vh − Vh )w, v)| ≤ |Vhλ,ij [, k] − Vhλ,ij [, k]| |wk | |v |
λ=0 j=1 i=1 τk ∈ωiλ τ ∈ωjλ
4 56 7
ωiλ ,ωjλ maximally admissible
L 4
4
λ λ
≤ c(η, ) ∆k |wk | ∆ |v | .
λ=0 j=1 i=1 τk ∈ωiλ τ ∈ωjλ
4 56 7
ωiλ ,ωjλ maximally admissible
Due to the assumption of the maximal admissibility of clusters ωiλ and ωjλ
each pair of boundary elements τk and τ appears maximal only once. Hence
we have
N
N
|((Vh − Vh )u, v)| ≤ c(η, ) ∆k |wk | ∆ |v | .
k=1 =1
By using the error estimate of Lemma 14.1, by applying the inverse inequality
in Sh0 (Γ ), and by an appropriate choice of the parameter η for the definition of
the near field and of the approximation order we now can prove the positive
definiteness of the approximate stiffness matrix Vh .
Theorem 14.2. Let all assumptions of Lemma 14.1 be satisfied. For an ap-
propriate choice of the parameter η and the approximate stiffness matrix Vh
is positive definite, i.e.
1 V
(Vh w, w) ≥ c wh 2H −1/2 (Γ ) for all wh ∈ Sh0 (Γ ).
2 1
Proof. By using Lemma 14.1, the H −1/2 (Γ )–ellipticity of the single layer po-
tential V , and the inverse inequality in Sh0 (Γ ) we obtain
Theorem 14.3. Let the parameter η and be chosen such that the approxi-
mate stiffness matrix Vh as defined in (14.15) is positive definite. The uniquely
determined solution w ∈ RN ↔ w h ∈ Sh0 (Γ ) of the perturbed linear system
Vh w
= f then satisfies the error estimate
c(η, ) h−1/2 wL2 (Γ ) .
h H −1/2 (Γ ) ≤ w − wh H −1/2 (Γ ) +
w − w
wh L2 (Γ ) ≤ wL2 (Γ ) + w − wh L2 (Γ ) ≤ c wL2 (Γ ) .
wh − w c(η, ) h−1/2 wL2 (Γ ) .
h H −1/2 (Γ ) ≤
The assigned error estimate now follows from the triangle inequality.
1
From the error estimate (14.3) we conclude, when assuming w ∈ Hpw (Γ ),
p 1
1 dn 1 p d
p+1
f (1) = f (0) + f (t)|t=0 + (1 − s) f (s)ds.
n=1
n! dtn p! dsp+1
0
14.2 Approximation of the Stiffness Matrix 337
Let yiλ be the center of the cluster ωiλ . For an arbitrary y ∈ ωiλ and t ∈ [0, 1]
let
f (t) := U ∗ (x, yiλ + t(y − yiλ )).
Then we have
2
d ∂ ∗
f (t) = (yj − yi,j
λ
) U (x, z)|z=yiλ +t(y−yiλ ) ,
dt j=1
∂z j
Thus, the Taylor expansion of the fundamental solution U ∗ (x, y) with respect
to the cluster center yiλ gives the representation
where
p
1
U∗ (x, y) = U ∗ (x, yiλ ) + (y − yiλ )α Dzα U ∗ (x, z)|z=yiλ (14.19)
α!
n=1 |α|=n
and
1
λ,j
fn,α (x) := Dzα U ∗ (x, z)|z=yiλ , (y − yiλ )α
λ,i
gn,α (y) :=
α!
for n = 1, . . . , p and |α| = n, we then obtain the representation (14.11). For
any n ∈ [1, p] there exist n + 1 multi–indices α ∈ N20 satisfying |α| = n. Then,
the number of terms in the series representation (14.11) is
p
1
= 1+ (n + 1) = (p + 1)(p + 2).
n=1
2
1
1
Rp (y, yiλ ) = (1 − s)p (y − yiλ )α Dzα U ∗ (x, z)|z=yiλ +s(y−yiλ ) ds.
p!
0 |α|=p+1
For this we first need to estimate certain derivatives of the fundamental solu-
tion U ∗ (x, y).
338 14 Fast Boundary Element Methods
Lemma 14.4. Let ωiλ and ωjλ be a pair of maximally admissible clusters. For
|α| = p ∈ N there holds the estimate
α ∗
Dy U (x, y) ≤ 1 3 (p − 1)!
p−1
for all (x, y) ∈ ωjλ × ωiλ .
2π |x − y|p
Proof. For the derivatives of the function f (x, y) = log |x − y| we first have
∂ yi − xi
f (x, y) = for i = 1, 2.
∂yi |x − y|2
and
∂2 (y1 − x1 )(y2 − x2 )
f (x, y) = −2 .
∂y1 ∂y2 |x − y|4
In general we find for |α| = ∈ N a representation of the form
(y − x)β
Dyα f (x, y) = aβ (14.20)
|x − y||β|+
|β|≤
A comparison with the first and second order derivatives of f (x, y) gives c1 = 1
and c2 = 3. Now, a general estimate of the constant c for ≥ 2 follows by
induction. From (14.20) we obtain for i = 1, 2 and j = i
∂ α ∂ (y − x)β
Dy f (x, y) = aβ
∂yi ∂yi |x − y||β|+
|β|≤
(yi − xi )βi −1 (yj − xj )βj
= aβ βi
|x − y||β|+
|β|≤
(yi − xi )βi +1 (yj − xj )βj
−(|β| + ) .
|x − y||β|++2
By using
we then obtain
14.2 Approximation of the Stiffness Matrix 339
∂ α 3 3c c+1
∂yi Dy f (x, y) ≤ |x − y|+1 aβ =
|x − y|+1
=
|x − y|+1
|β|≤
and therefore
c+1 = 3c = 3 ! .
In particular for = p this gives the assertion.
By using Lemma 14.4 we now can derive an estimate of the remainder
Rp (x, y) of the Taylor series approximation (14.19).
Lemma 14.5. Let ωiλ and ωjλ be a pair of maximally admissible clusters. For
the approximation U∗ (x, y) of the fundamental solution U ∗ (x, y) as defined in
(14.19) there holds the error estimate
p+1
∗
U (x, y) − U∗ (x, y) ≤ 3p 1 for all (x, y) ∈ ωjλ × ωiλ .
η
Proof. By applying Lemma 14.4 for (x, y) ∈ ωjλ × ωiλ and by using the admis-
sibility condition (14.8) we obtain
∗
U (x, y) − U∗ (x, y) = Rp (y, yiλ )
1
1
≤ |y − yiλ |p+1 max |Dzα U ∗ (x, z)z=ȳ | (1 − s)p ds
p! ȳ∈ωiλ
|α|=p+1 0
1 3p p!
= |y − yiλ |p+1 max
(p + 1)! ȳ∈ωiλ |x − ȳ|p+1
|α|=p+1
p+1
[diam ωiλ ]p+1 1
≤ 3p ≤ 3p .
[dist(ωiλ , ωjλ )]p+1 η
) = |y − yiλ | eiϕ(y−yi ) .
λ
z0 := (y1 − yi,1
λ
) + i(y2 − yi,2
λ
Since the clusters ωiλ and ωjλ are assumed to be admissible, it follows that
|z0 | |y − yiλ | diam ωiλ 1
= ≤ ≤ < 1.
|z| |x − yi |
λ dist(ωiλ , ωjλ ) η
Hence we can apply the series representation of the logarithm,
1 1 1 9 z0 :
− log(z − z0 ) = − log z − log 1 −
2π 2π 2π z
1 z 0 :n
9
∞
1 1
=− log z + ,
2π 2π n=1 n z
and for p ∈ N we can define an approximation
& '
1 1 9 z 0 :n
p
∗ 1
U (x, y) := Re − log z + (14.21)
2π 2π n=1 n z
of the fundamental solution U ∗ (x, y). By using
9 z :n |y − yiλ |n inϕ(y−yiλ ) −inϕ(x−yiλ )
0
= z0n z −n = e e
z |x − yiλ |n
we then obtain the representation
1
U∗ (x, y) = − log |x − yiλ |
2π
1 1
p
cos nϕ(x − yiλ )
+ |y − yiλ |n cos nϕ(y − yiλ )
2π n=1 n |x − yiλ |n
1 1
p
sin nϕ(x − yiλ )
+ |y − yiλ |n sin nϕ(y − yiλ ) .
2π n=1 n |x − yiλ |n
14.2 Approximation of the Stiffness Matrix 341
By introducing
f0λ,j (x) := U ∗ (x, yiλ ), g0λ,i (y) := 1
and
as well as
Lemma 14.6. Let ωiλ and ωjλ be a pair of admissible clusters. For the approx-
imation U (x, y) of the fundamental solution U ∗ (x, y) as defined in (14.21)
there holds the error estimate
p
∗ ∗ 1 1 1 1
|U (x, y) − U (x, y)| ≤ for all (x, y) ∈ ωjλ × ωiλ .
2π p + 1 η − 1 η
1 1 9 z0 : n
∞
1 1
− log(z − z0 ) = − log z +
2π 2π 2π n=1 n z
To ensure the asymptotically optimal error estimate (14.3) the related condi-
tion (14.18) now reads
p
1 1 1 1
≤ c h2 .
2π p + 1 η − 1 η
p = O(log2 N ) .
342 14 Fast Boundary Element Methods
The total amount of work (14.16) to store Vh and to realize a matrix by vector
multiplication with the approximated stiffness matrix Vh is then proportional
to
N (log2 N )2 .
Note that both the approximation (14.19) based on the Taylor expansion
as well as the series expansion (14.21) define nonsymmetric approximations
U∗ (x, y) of the fundamental solution U ∗ (x, y), and therefore this results in
a nonsymmetric approximated stiffness matrix Vh . Hence we aim to derive a
symmetric approximation U∗ (x, y). For this we first consider the representa-
tion (14.21),
& '
1 1 p
1 9 z :n
0
U∗ (x, y) = Re − log z +
2π 2π n=1 n z
where
z = |x − yiλ |eiϕ(x−yi ) , z0 = |y − yiλ |eiϕ(y−yi ) .
λ λ
1 1 9 z1 : n
∞
1 1
− log z = − log w + .
2π 2π 2π n=1 n w
Lemma 14.7. Let w, z1 ∈ C satisfying |z1 | < |w|. For n ∈ N then there holds
∞
1 m+n−1 z1m
= .
(w − z1 )n n−1 wm+n
m=0
1 9 z1 :m
∞
1 1 1
= =
w − z1 w 1 − z1 w m=0 w
w
and therefore the assertion in the case n = 1. For n > 1 we have
14.2 Approximation of the Stiffness Matrix 343
1 1 dn−1 1
=
(w − z1 )n (n − 1)! dz1n−1 w − z1
# $
dn−1 1 9 z1 :m
∞
1
=
(n − 1)! dz1n−1 w m=0 w
∞
1 1 m! z1m−n+1
=
(n − 1)! w m=n−1
(m − n + 1)! wm
∞
1 1 (m + n − 1)! z1m
=
(n − 1)! w m=0 m! wm+n−1
∞
m+n−1 z1m
= .
n−1 wm+n
m=0
Lemma 14.8. Let ωiλ and ωjλ be a pair of admissible clusters. For the approx-
imation U ∗ (x, y) of the fundamental solution U ∗ (x, y) as defined in (14.22)
there holds the error estimate
p
∗ ∗ (x, y) ≤ 1 1 1 1 1
U (x, y) − U 2 +
2π p + 1 η − 1 η−1 η
rk−1 (x, yk )rk−1 (xk , y)
U∗ (x, y) = s (x, y) = (14.25)
rk−1 (xk , yk )
k=1
Lemma 14.9. For 0 ≤ k ≤ and for all (x, y) ∈ ωjλ × ωiλ there holds
as well as
rk (xj , y) = 0 for all 1 ≤ j ≤ k. (14.28)
Proof. By taking the sum of the recursions (14.23) and (14.24) we first have
This means that the approximations sk (x, y) interpolate the fundamental so-
lution U ∗ (x, y) at the interpolation nodes (xj , yi ) for i, j = 1, . . . , k.
To analyze the approximations U∗ (x, y) as defined via (14.23) and (14.24)
we consider a sequence of matrices
,
k ,
k
det Mk = r0 (x1 , y1 ) · · · rk−1 (xk , yk ) = ri−1 (xi , yi ) = αi . (14.30)
i=1 i=1
Assume that (14.30) holds for Mk . The matrix Mk+1 allows the representation
⎛ ∗ ⎞
U (x1 , y1 ) · · · U ∗ (x1 , yk ) U ∗ (x1 , yk+1 )
⎜ .. .. .. ⎟
⎜ . . . ⎟
Mk+1 = ⎜ ⎟.
⎝ U ∗ (xk , y1 ) · · · U ∗ (xk , yk ) U ∗ (xk , yk+1 ) ⎠
U ∗ (xk+1 , y1 ) · · · U ∗ (xk+1 , yk ) U ∗ (xk+1 , yk+1 )
r0 (x, yi ) = U ∗ (x, yi ),
r0 (x, y1 )r0 (x1 , yi )
r1 (x, yi ) = r0 (x, yi ) −
r0 (x1 , y1 )
r0 (x1 , yi ) ∗
= U ∗ (x, yi ) − U (x, y1 )
r0 (x1 , y1 )
and therefore
r0 (x1 , yi )
r1 (x, yi ) = U ∗ (x, yi ) − α11 (yi ) U ∗ (x, y1 ), α11 (yi ) := .
r0 (x1 , y1 )
k
rk (x, yi ) = U ∗ (x, yi ) − αjk (y)U ∗ (x, yj ) (14.31)
j=1
By using the recursion (14.24) and inserting twice the assumption of the
induction this gives
14.2 Approximation of the Stiffness Matrix 347
k+1
∗
= U (x, yi ) − αjk+1 (y)U ∗ (x, yj ) .
j=1
By
-k+1 [j, i] := Mk+1 [j, i]
M for i = 1, . . . , k, j = 1, . . . , k + 1
and
k
-k+1 [j, k + 1] = Mk+1 [j, k + 1] −
M α k (y)Mk+1 [j, ] for j = 1, . . . , k + 1
=1
we define a transformed matrix M -k+1 satisfying det M -k+1 = det Mk+1 . Insert-
ing the definition of Mk+1 [j, ·] this gives for j = 1, . . . , k + 1, due to (14.31),
k
-k+1 [j, k + 1] = U ∗ (xj , yk+1 ) −
M α k (y)U ∗ (xj , y ) = rk (xj , yk+1 ).
=1
k
1
sk (x, y) = ri−1 (x, yi )ri−1 (xi , y) .
α
i=1 i
From (14.31) we find for the residual ri−1 (x, yi ) the representation
i−1
∗
ri−1 (x, yi ) = U (x, yi ) − α i−1 (y)U ∗ (x, y ).
=1
i−1
ri−1 (xi , y) = U ∗ (xi , y) − β i−1 (x)U ∗ (x , y).
=1
Mk = Mk Ak Mk
Lk (y ) = δk for k, = 1, . . . , .
14.2 Approximation of the Stiffness Matrix 349
Let ⎛ ⎞
U ∗ (x1 , y1 ) · · · · · · U ∗ (x1 , y )
⎜ .. .. ⎟
⎜. . ⎟
⎜ ∗ ⎟
⎜ U (x −1 , y1 ) ······ U (x −1 , y ) ⎟
∗
⎜ ∗ ⎟
M, (x) := ⎜
⎜ U ∗ (x, y1 ) ······ U ∗ (x, y ) ⎟.
⎟
⎜ U (x +1 , y1 ) ······ U (x +1 , y ) ⎟
∗
⎜ ⎟
⎜. .. ⎟
⎝ .. . ⎠
U ∗ (x , y1 ) ······ ∗
U (x , y )
Lemma 14.12. For (x, y) ∈ ωjλ × ωiκ let r (x, y) be the residual of the ap-
proximation U∗ (x, y). Then there holds
det M,k (x)
r (x, y) = E (x, y) − E (xk , y).
det M
k=1
Proof. By using Lemma 14.9 and the matrix representation of s (x, y), see
Lemma 14.11, we first have
Due to
& '
∗ ∗ ∗
(U (xk , y))k=1, − M (Lk (y))k=1, = U (xk , y) − U (xk , y )L (y)
=1 k=1,
= (E (xk , y))k=1,
we then conclude
r (x, y) = E (x, y) − (U ∗ (x, yk ))k=1, M−1 (E (xk , y))k=1,
9 :
= E (x, y) − M− (U ∗ (x, y )) =1, (E (xk , y))k=1, .
350 14 Fast Boundary Element Methods
and thus
ei = M− (U ∗ (xi , yk ))k=1, .
Hence we conclude
M− M,
(x) =
9 :
= M− (U ∗ (xi , yk ))k=1,;i=1, −1 , (U ∗ (x, yk ))k=1, , (U ∗ (xi , yk ))k=1,;i= +1,
9 :
= e1 , . . . , e −1 , M− (U ∗ (x, yk ))k=1, , e +1 , . . . , e ,
and with
9 : det M,k (x)
M− (U ∗ (x, y ) =1, = det M− M,k
(x) =
k det M
Corollary 14.13. Let the interpolation nodes (xk , yk ) ∈ ωjλ × ωiλ be chosen
such that
|det M, (x)| ≤ |det M | (14.32)
is satisfied for all = 1, . . . , and for all x ∈ ωjλ . The residual r (x, y) then
satisfies the estimate
The criteria (14.32) to define the interpolation nodes (xk , yk ) ∈ ωjλ ∈ ωiλ seems
not be very suitable for a practical realization. Hence we finally consider an
alternative choice.
Lemma 14.14. For k = 1, . . . , let the nodal pairs (xk , yk ) ∈ ωjλ ∈ ωiλ be
chosen such that
Proof. As in the proof of Lemma 14.10 we have for det Mk, (x) and 1 ≤ < k
the recursion
det Mk, (x) = rk−1 (xk , yk )det Mk−1, (x) − rk−1 (x, yk )det Mk−1, (xk ),
14.3 Wavelets 351
or,
det M1,1 (x) = r0 (x, y1 ), det Mk,k (x) = rk−1 (x, yk )det Mk−1
for k = 2, 3, . . . , , as well as
det Mk, (x) det Mk−1, (x) rk−1 (x, yk ) det Mk−1, (x)
= −
det Mk det Mk−1 rk−1 (xk , yk ) det Mk−1
and therefore
|det Mk, (x)| |det Mk−1, (x)|
sup ≤ 2 sup ,
x∈ωjλ |det Mk | x∈ωjλ |det Mk−1 |
or,
det Mk,k (x)
= rk−1 (x, yk ) ≤ 1 .
det Mk rk−1 (xk , yk )
Corollary 14.15. For k = 1, . . . , let the nodal pairs (xk , yk ) ∈ ωjλ × ωiλ be
chosen such that assumption (14.33) is satisfied. Then there holds
14.3 Wavelets
In this subsection we introduce wavelets as hierarchical basis functions to
be used in the Galerkin discretization of the single layer potential V . As in
standard boundary element methods this leads to a dense stiffness matrix,
but by neglecting small matrix entries one can define a sparse approximation.
352 14 Fast Boundary Element Methods
This reduces both the amount of storage, and amount to realize a matrix by
vector multiplication.
Without loss of generality we assume that the Lipschitz boundary Γ = ∂Ω
of a bounded domain Ω ⊂ R2 is given via a parametrization Γ = χ(Q) with
respect to a parameter domain Q = [0, 1] where we assume that the Jacobian
|χ̇(ξ)| is constant for all ξ ∈ Q. Moreover, we extend the parametrization
Γ = χ([0, 1]) periodically onto R. Hence we can assume the estimates
with some positive constants cχ1 and cχ2 . In the case of a piecewise smooth
Lipschitz boundary Γ all following considerations have to be transfered to the
non–periodic parametrizations describing the parts Γj satisfying |χ̇j | = cj .
For j ∈ N we consider a decomposition of the parameter domain Q = [0, 1]
into Nj = 2j finite elements q j of mesh size |q j | = 2−j , = 1, . . . , Nj ,
2
*
j
where
VL = W0 ⊕ W1 ⊕ · · · ⊕ WL .
Due to W0 = V0 we have
14.3 Wavelets 353
Wj = span{ψ j }2 =1
j−1
for j = 1, . . . , L.
ψ11 (ξ) = a1 ϕ
11 (ξ) + a2 ϕ
12 (ξ) for ξ ∈ Q = [0, 1]
which holds for piecewise constant wavelets ψ j . The basis functions (14.35)
as constructed above are also denoted as Haar wavelets, see Fig. 14.4.
354 14 Fast Boundary Element Methods
1 ψ10
-
0 1
−1 ψ11
ψ12 ψ22
<Nj j
The global mesh size of the boundary element mesh ΓNj = =1 τ is then
given as hj = |χ̇| 2−j . Moreover, we can lift both the piecewise constant basis
j ∈ Vj as well as the wavelets ψ j ∈ Wj on the boundary Γ = ∂Ω,
functions ϕ
for x = χ(ξ) ∈ Γ we have
ϕj (x) := ϕ
j (ξ), ψ j (x) := ψ j (ξ) for ξ ∈ Q = [0, 1].
VL = Sh0L (Γ ) = span{ϕL j
} =1 = span{ψ } =1,...,max{1,2j−1 },j=1,...,L ,
NL
L max{1,2 }
i−1
whL = wki ψki ∈ VL . (14.37)
i=0 k=1
Due to
ψki , ψ j L2 (Ω) = ψki (x)ψ j (x)dsx
Γ
|χ̇| for i = j, k = ,
= |χ̇| ψki (ξ)ψ j (ξ)dξ =
0 elsewhere
Q
the orthogonality of the trial spaces Wki in the parameter domain is transfered
to the trial spaces Wki which are defined with respect to the boundary element
mesh.
By using Remark 13.21 we can derive spectrally equivalent norm represen-
tations by means of the multilevel representation (14.37) of a given function
whL ∈ VL .
max{1,2i−1 }
L i 2
whL 2L,s = 22si wk .
i=0 k=1
L
B whL , whL L2 (Γ ) =
s
h−2s
i (Qi − Qi−1 )whL 2L2 (Γ ) ,
i=0
Qi whL , vhi L2 (Γ ) = whL , vhi L2 (Γ ) for all vhi ∈ Vi .
max{1,2i−1 }
(Qi − Qi−1 )whL = wki ψki
k=1
and therefore
= =2
=max{1,2i−1 } =
L
= =
−2s = i i=
B s whL , whL L2 (Γ ) = hi = wk ψk =
i=0 = k=1 =
L2 (Γ )
max{1,2i−1 }
L i 2
= |χ̇| h−2s wk .
i
i=0 k=1
By inserting the mesh sizes hi = |χ̇| 2−i this concludes the proof.
By using the representation (14.37) the variational problem (14.1) is equiv-
alent to
L max{1,2 }
i−1
wki V ψki , ψ j Γ = g, ψ j Γ for all ψ j ∈ VL . (14.38)
i=0 k=1
and 9 :
dij i j
k := dist Sk , S = min |x − y|
(x,y)∈Ski ×Sj
describes the distance between the supports of the basis functions ψki and ψ j ,
respectively.
Lemma 14.17. Assume that for i, j ≥ 2 the condition dij k > 0 is satisfied.
Then there holds the estimate
|χ̇|4 −3(i+j)/2 9 ij :−2
|V L [(, j), (k, i)]| ≤ 2 dk .
2π
14.3 Wavelets 357
Due to the moment condition (14.36) we can replace the kernel function k(s, t)
by r(s, t) := k(s, t) − P1 (s) − P2 (t) where P1 (s) and P2 (t) correspond to
the first terms of the Taylor expansion of k(s, t), i.e. r(s, t) corresponds to
the remainder of the Taylor expansion. The Taylor expansion of the kernel
function k(s, t) with respect to s0 = 12 gives
1 1 ∂
k(s, t) = k ,t + s − k(s, t)
2 2 ∂s s=s̄
Hence we obtain
|χ̇|2 −(i+j)/2
V [(, j), (k, i)] = −
L
2 ψ%11 (ξ(t)) k(s, t)ψ%11 (η(s)) ds dt
π
Q Q
2
|χ̇| −(i+j)/2 1 1
=− 2 ψ%11 (ξ(t)) s− t− ·
π 2 2
Q Q
∂2
· k(s, t) ψ%1 (η(s)) ds dt
1
∂s∂t (s,t)=(s̄,t̄)
and therefore
2
L 2 ∂
V [(, j), (k, i)] ≤ |χ̇| 2−(i+j)/2 1
max k(s, t).
π 16
(s,t)∈Q×Q ∂s∂t
2
∂ ∂ ∂
log |χ(η) − χ(ξ)| = log |y − x(ξ)||y=χ(η) χj (η),
∂η i=1
∂y i ∂η
as well as
2
∂2 ∂2 ∂ ∂
log |χ(η)−χ(ξ)| = log |y−x||y=χ(η),x=χ(ξ) χi (η) χi (ξ) .
∂η∂ξ i,j=1
∂y i ∂xj ∂η ∂ξ
By using
∂2 1 (xi − yi )2
log |x − y| = − 2
+2 ,
∂xi ∂yi |x − y| |x − y|4
∂2 (x1 − y1 )(x2 − y2 )
log |x − y| = 2
∂x1 ∂y2 |x − y|4
14.3 Wavelets 359
we finally obtain
2
∂ 2
max k(s, t) ≤ 22−(i+j) |χ̇|2 max .
(s,t)∈Q×Q ∂s∂t (x,y)∈Ski ×Sj |x − y|2
The estimate of Lemma 14.17 describes the decay of the matrix entries
V L [(, j), (k, i)] when considering wavelets ψki and ψ j with supports Ski and S j
which are far to each other. By defining an appropriate compression parameter
we therefore can characterize matrix entries V L [(, j), (k, i)] which can be
neglected when computing the stiffness matrix V L . For real valued parameter
α, κ ≥ 1 we first define a symmetric parameter matrix by
τij := α 2κL−i−j .
For an arbitrary but fixed = 1, . . . , 2j−1 we first determine all basis functions
ψki where the supports Ski and S j do not overlap, i.e.
Lemma 14.19. For i, j = 2, . . . , L let VijL be the exact Galerkin stiffness ma-
trix of the single layer potential V with respect to the trial spaces Wi and Wj .
Let VijL be the approximation as defined in (14.39). Then there hold the error
estimates
2
i−1
L
VijL − VijL ∞ = maxj−1 V [(, j), (k, i)] − V L [(, j), (k, i)]
=1,...,2
k=1
2
i−1
L
= maxj−1 V [(, j), (k, i)]
=1,...,2
k=1
dij
k >τij
2
i−1
−3(i+j)/2 −2
≤ c2 max (dij
k ) .
=1,...,2j−1
k=1
dij
k >τij
VijL − VijL ∞ ≤
c2 −3(i+j)/2
max (dist(Ski , S j ))−2 .
=1,...,2j−1
k=1
cχ i j
1 dist(Sk ,S )>τij
For an arbitrary but fixed = 1, . . . , 2j−1 the sum can be further estimated
by
2
i−1
−2
(dist(Ski , S j ))−2 ≤ 2 2(k − 1)2−i − 22−j
k=1 k>1+ α
χ 2κL−j + 2i−j
2c1
cχ i j
1 dist(Sk ,S )>τij
1
= 22i−1 .
n2
n> α
χ 2κL−j
2c1
∞
∞ ∞
1 1 1 1 1 1 1 2 4cχ1 j−κL
= + ≤ + dx = + ≤ ≤ 2
n=n
n2 n21 n=n +1 n2 n21 x2 n21 n1 n1 α
1 1 x=n1
362 14 Fast Boundary Element Methods
which gives immediately the first estimate. The second estimate follows in the
same way.
By applying Lemma 13.17 (Schur Lemma) we can now estimate the error
of the approximation V L .
where
⎧ κ+σ1 +σ2
⎪
⎪ hL for σ1 , σ2 ∈ (− 12 , 0),
⎪
⎨ hκ | ln h |
L L for σ1 = σ2 = 0,
γ(hL , σ1 , σ2 ) :=
⎪
⎪ h κ
for σ1 , σ2 ∈ (0, 12 ),
⎪ L
⎩ κ+σ2
hL for σ1 ∈ (0, 12 ), σ2 ∈ (− 12 , 0).
≤ A2 wh L,σ1 vh L,σ2
≤ A2 wh H σ1 (Γ ) vh H σ2 (Γ )
By using (13.17) we now can estimate the spectral norm A2 for an arbitrary
s as
14.3 Wavelets 363
L 2
i−1
L 2
i−1
L
V [(, j), (k, i)]− V [(, j), (k, i)]
s(i−j) −σ1 i−σ2 j L
= sup 2 2 sup
j=2,...,L i=2 =1,...,2j−1 k=1
L
= sup 2s(i−j) 2−σ1 i−σ2 j VijL − VijL ∞
j=2,...,L i=2
L
≤c sup 2s(i−j) 2−σ1 i−σ2 j 2−(i+j)/2 2−j (τij )−1
j=2,...,L i=2
L
1 i+j−κL
=c sup 2s(i−j) 2−σ1 i−σ2 j 2−(i+j)/2 2−j 2
j=2,...,L i=2 α
1
L
1
c 2−κL
= sup 2j(−s− 2 −1+1−σ2 ) 2i(s− 2 +1−σ1 )
j=2,...,L i=2
L
c 2−κL
= sup 2−σ2 j 2−σ1 i
j=2,...,L i=2
For this we first need to establish the positive definiteness of the approximated
stiffness matrix V L .
Theorem 14.21. For κ > 1 and for a sufficient small global boundary ele-
ment mesh size hL the approximated stiffness matrix V L is positive definite,
i.e.
1
(V L w, w) ≥ cV1 whL 2H −1/2 (Γ )
2
is satisfied for all whL ∈ VL ↔ w ∈ RN .
Now, if
chκ−1
L ≤ 12 cV1 is satisfied the assertion follows.
Instead of the linear system V L w = f which corresponds to the variational
problem (14.38) we now have to solve the perturbed linear system V L w = f
where w ∈ RN ↔ w hL ∈ VL defines the associated approximate solution.
1
Theorem 14.22. Let w ∈ Hpw (Γ ) be the unique solution of the boundary
integral equation V w = g. For the approximate solution w hL ∈ VL ↔ w ∈ RN
of the perturbed linear system V L w
= f there holds the error estimate
3/2 κ−1/2
w − w
hL H −1/2 (Γ ) ≤ c1 hL wHpw
1 (Γ ) + c2 h
L wH 1/2 (Γ ) .
(V L w − V L w,
v) = 0 for all v ∈ RN .
1 V
c whL − w
hL H −1/2 (Γ ) ≤ c hκ+σ 2
whL H σ1 (Γ ) whL − w
hL H σ2 (Γ ) .
2 1 L
Hence we have
1 V κ−1/2
c whL − w
hL H −1/2 (Γ ) ≤ c hL wH 1/2 (Γ ) .
2 1
Now the assertion follows from applying the triangle inequality
Remark 14.23. The error estimate of Theorem 14.22 is not optimal with re-
1
spect to the regularity of the solution w ∈ Hpw (Γ ). Since Lemma 14.16 is
1 1 1
only valid for s ∈ (− 2 , 2 ) the higher regularity w ∈ Hpw (Γ ) is not recognized
in the error estimate. Formally, this yields the error estimate
2 3
3/2
w − whL H −1/2 (Γ ) ≤ c hL + hκL wHpw 1 (Γ ) .
14.4 Exercises
14.1 Consider the finite element stiffness matrix of Exercise 11.1 for h = 1/9,
⎛ ⎞
2 −1
⎜ −1 2 −1 ⎟
⎜ ⎟
⎜ −1 2 −1 ⎟
⎜ ⎟
⎜ −1 2 −1 ⎟
Kh = 9 ⎜⎜
⎟.
⎟
⎜ −1 2 −1 ⎟
⎜ −1 2 −1 ⎟
⎜ ⎟
⎝ −1 2 −1 ⎠
−1 2
is given by
1
u(x) = (N f )(x) = G(x, y)f (y)dy for x ∈ (0, 1)
0
where G(x, y) is the associated Green function, cf. Exercise 5.2. Discuss the
Galerkin discretization of N f when using piecewise linear continuous basis
functions with respect to a uniform decomposition of (0, 1).
15
Domain Decomposition Methods
*
p
Ω = Ωi, Ωi ∩ Ωj = ∅ for i = j. (15.2)
i=1
For an approximate solution of the boundary value problem (15.1) we will use
a boundary element method within the subdomains Ω1 , . . . , Ωq while for the
remaining subdomains Ωq+1 , . . . , Ωp a finite element method will be applied.
368 15 Domain Decomposition Methods
Ω3 Ω4
ΓS
Ω1 Ω2
p
αi ∇u(x)∇v(x)dx = 0 for all v ∈ H01 (Ω).
i=1 Ωi
The application of Green’s first formula (1.5) with respect to the subdomains
Ωi for i = 1, . . . , q ≤ p results in a variational problem to find u ∈ H 1 (Ω)
with γ0int u = g such that
q
p
αi int u(x)γ int v(x)ds +
γ1,i αi ∇u(x)∇v(x)dx = 0 (15.5)
0,i x
i=1 ∂Ωi i=q+1 Ωi
Inserting the second equation of (15.6) into the variational formulation (15.5)
this results in the variational problem to find u ∈ H 1 (Ω) with γ0int u = g and
int u ∈ H −1/2 (Γ ) for i = 1, . . . , q such that
γ1,i i
15. Domain Decomposition Methods 369
q p
int u + ( 1 I + K )γ int u, γ int v +
αi Di γ0,i α ∇u(x)∇v(x)dx = 0
i 1,i 0,i Γi i
i=1
2 i=q+1 Ωi
1
int u − ( I + K )γ int u, τ
αi Vi γ1,i i 0,i i Γi = 0
2
a(u, γ int
1
u; v, τ ) = 0 (15.7)
The boundedness of the bilinear form a(·, ·) follows from the boundedness of
all local boundary integral operators, and from the boundedness of the local
Dirichlet forms.
For arbitrary (v, τ ) ∈ X we have
370 15 Domain Decomposition Methods
q 2 3
p
a(v, τ ; v, τ ) = int v, γ int v
αi Vi τi , τi Γi + Di γ0,i 0,i Γi + αi ∇v2L2 (Ωi )
i=1 i=q+1
q
V 2 3
≥ min αi c1,i , αi c1,i , αi
D
τi 2H −1/2 (Γi ) |γ0,i
int u|2
H 1/2 (Γi )
i=1,p
i=1
⎫
p ⎬
+ ∇v2L2 (Ωi ) .
⎭
i=q+1
Due to v ∈ H01 (Ω) we therefore conclude the X–ellipticity of the bilinear form
a(·, ·) and thus the unique solvability of the variational problem (15.7).
Let
Xh := Sh1 (Ω) × Sh0 (Γ1 ) × · · · × Sh0 (Γq ) ⊂ X
be a conforming trial space of piecewise linear basis functions to approximate
the potential u ∈ H01 (Ω) and of piecewise constant basis functions to approx-
imate the local Neumann data γ1,i int u ∈ H −1/2 (Γ ), i = 1, . . . , q. All degrees of
i
freedom of the trial space Sh1 (Ω) ⊂ H01 (Ω) are depicted in Fig. 15.2.
t
t
t
BEM
t
t
t
t
t t t t t t t t t t t t t t t
t
t
t
t BEM
t
t
t
Fig. 15.2. Degrees of freedom of the trial space Sh1 (Ω) ⊂ H01 (Ω).
The global trial space Sh1 (Ω) ⊂ H01 (Ω) is decomposed into local trial spaces
which is defined with respect to the skeleton ΓS . All global degrees of freedom
are characterized in Fig. 15.2 by • while all local degrees of freedom correspond
to . From the decomposition
*
p
Sh1 (Ω) = Sh1 (ΓS ) ∪ Sh1 (Ωi )
i=q+1
it follows that a function uh ∈ Sh1 (Ω) ∩ H01 (Ω) allows the representation
MS
p
Mi
uh (x) = uS,k ϕ1S,k (x) + ui,k ϕ1i,k (x).
k=1 i=q+1 k=1
a(u0,h + ug , th ; vh , τ h ) = 0 (15.8)
is satisfied for all vh ∈ Sh1 (Ω) ∩ H01 (Ω) and τi,h ∈ Sh0 (Γi ), i = 1, . . . , q.
By applying Theorem 8.1 (Cea’s Lemma) the Galerkin variational formu-
lation (15.8) has a unique solution which satisfies the a priori error estimate
with
Vh,i = αi Vi ϕ0i,k , ϕ0i, Γi for k, = 1, . . . , Ni , i = 1, . . . , q,
and
Ah,i = αi ∇ϕ1i,k (x)∇ϕ1i, (x)dx for k, = 1, . . . , Mi , i = q + 1, . . . , p.
Ωi
The linear system (15.9) corresponds to the general system (13.20), hence we
can apply all iterative methods of Chapter 13.3 to solve (15.9). In particular,
when eliminating the local degrees of freedom t and uL we obtain the Schur
complement system
1 1
Dh + ( Mh + Kh )Vh−1 ( Mh + Kh ) + ASS − ALS A−1 LL SL uS = f
A
2 2
(15.11)
where the modified right hand side is given by
1
f := f S − ( Mh + Kh )Vh−1 f B − ALS A−1
LL f L .
2
Due to (15.10) the inversion of the local stiffness matrices Vh and ALL can
be done in parallel. This corresponds to the solution of local Dirichlet bound-
ary value problems. In general we have to use local preconditioners for the
local stiffness matrices Vh,i and Ah,i . For the solution of the global Schur
complement system (15.11) where the system matrix Sh is symmetric and
positive definite, we can use a preconditioned conjugate gradient scheme. The
definition of an appropriate preconditioning matrix is then based on spectral
equivalence inequalities of the corresponding Schur complement matrices,
1 1
ShBEM := Dh + ( Mh + Kh )Vh−1 ( Mh + Kh ),
2 2
and
ShFEM := ASS − ASL A−1
LL ALS ,
1. Adams, R. A.: Sobolev Spaces. Academic Press, New York, London, 1975.
2. Ainsworth, M., Oden, J. T.: A Posteriori Error Estimation in Finite Element
Analysis. John Wiley & Sons, New York, 2000.
3. Atkinson, K. E.: The Numerical Solution of Integral Equations of the Second
Kind. Cambridge University Press, 1997.
4. Axelsson, O.: Iterative Solution Methods. Cambridge University Press,
Cambridge, 1994.
5. Axelsson, O., Barker, V. A.: Finite Element Solution of Boundary Value Prob-
lems: Theory and Computation. Academic Press, Orlando, 1984.
6. Aziz, A., Babus̆ka, I.: On the angle condition in the finite element method.
SIAM J. Numer. Anal. 13 (1976) 214–226.
7. Babus̆ka, I.: The finite element method with Lagrangian multipliers. Numer.
Math. 20 (1973) 179–192.
8. Babuška, I., Strouboulis, T.: The Finite Element Method and its Reliability.
The Clarendon Press, New York, 2001.
9. Banerjee, P. K.: The Boundary Element Methods in Engineering. McGraw,
London, 1994.
10. Bangerth, W., Rannacher, R.: Adaptive Finite Element Methods for Differential
Equations. Birkhäuser, Basel, 2003.
11. Barrett, R. et al.: Templates for the Solution of Linear Systems: Building Blocks
for Iterative Methods. SIAM, Philadelphia, 1993.
12. Bebendorf, M.: Effiziente numerische Lösung von Randintegralgleichungen unter
Verwendung von Niedrigrang–Matrizen. Dissertation, Universität des Saarlan-
des, Saarbrücken, 2000.
13. Bebendorf, M.: Approximation of boundary element matrices. Numer. Math. 86
(2000) 565–589.
14. Bebendorf, M., Rjasanow, S.: Adaptive low–rank approximation of collocation
matrices. Computing 70 (2003) 1–24.
15. Beer, G.: Programming the Boundary Element Method. John Wiley & Sons,
Chichester, 2001.
16. Bergh, J., Löfström, J.: Interpolation Spaces. An Introduction. Springer, Berlin,
New York, 1976.
17. Bespalov, A., Heuer, N.: The p–version of the boundary element method for
hypersingular operators on piecewise plane open surfaces. Numer. Math. 100
(2005) 185–209.
376 References
18. Bjørstad, P., Gropp, W., Smith, B.: Domain Decomposition. Parallel Multilevel
Methods for Elliptic Partial Differential Equations. Cambridge University Press,
1996.
19. Bonnet, M.: Boundary Integral Equation Methods for Solids and Fluids. John
Wiley & Sons, Chichester, 1999.
20. Bowman, F.: Introduction to Bessel Functions. Dover, New York, 1958.
21. Braess, D.: Finite Elemente. Springer, Berlin, 1991.
22. Braess, D.: Finite Elements: Theory, Fast Solvers and Applications in Solid
Mechanics. Cambridge University Press, 1997.
23. Brakhage, H., Werner, P.: Über das Dirichletsche Aussenraumproblem für die
Helmholtzsche Schwingungsgleichung. Arch. Math. 16 (1965) 325–329.
24. Bramble, J. H.: The Lagrange multiplier method for Dirichlet’s problem. Math.
Comp. 37 (1981) 1–11.
25. Bramble, J. H.: Multigrid Methods. Pitman Research Notes in Mathematics
Series, vol. 294, Longman, Harlow, 1993.
26. Bramble, J. H., Pasciak, J. E.: A preconditioning technique for indefinite sys-
tems resulting from mixed approximations of elliptic problems. Math. Comp.
50 (1988) 1–17.
27. Bramble, J. H., Pasciak, J. E., Steinbach, O.: On the stability of the L2 projec-
tion in H 1 (Ω). Math. Comp. 71 (2002) 147–156.
28. Bramble, J. H., Pasciak, J. E., Xu, J.: Parallel multilevel preconditioners. Math.
Comp. 55 (1990) 1–22.
29. Bramble, J. H., Zlamal, M.: Triangular elements in the finite element method.
Math. Comp. 24 (1970) 809–820.
30. Brebbia, C. A., Telles, J. C. F., Wrobel, L. C.: Boundary Element Techniques:
Theory and Applications in Engineering. Springer, Berlin, 1984.
31. Brenner, S., Scott, R. L.: The Mathematical Theory of Finite Element Methods.
Springer, New York, 1994.
32. Breuer, J.: Wavelet–Approximation der symmetrischen Variationsformulierung
von Randintegralgleichungen. Diplomarbeit, Mathematisches Institut A, Uni-
versität Stuttgart, 2001.
33. Buffa, A., Hiptmair, R.: Regularized combined field integral equations. Numer.
Math. 100 (2005) 1–19.
34. Buffa, A., Sauter, S.: Stabilisation of the acoustic single layer potential on non–
smooth domains. SIAM J. Sci. Comput. 28 (2003) 1974–1999.
35. Carstensen, C.: A unifying theory of a posteriori finite element error control.
Numer. Math. 100 (2005) 617–637.
36. Carstensen, C., Kuhn, M., Langer, U.: Fast parallel solvers for symmetric bound-
ary element domain decomposition methods. Numer. Math. 79 (1998) 321–347.
37. Carstensen, C., Stephan, E. P.: A posteriori error estimates for boundary ele-
ment methods. Math. Comp. 64 (1995) 483–500.
38. Carstensen, C., Stephan, E. P.: Adaptive boundary element methods for some
first kind integral equations. SIAM J. Numer. Anal. 33 (1996) 2166–2183.
39. Chen, G., Zhou, J.: Boundary Element Methods. Academic Press, New York,
1992.
40. Cheng, A. H.–D., Cheng, D. T.: Heritage and early history of the boundary
element method. Engrg. Anal. Boundary Elements 29 (2005) 268–302.
41. Ciarlet, P. G.: The Finite Element Method for Elliptic Problems. North–
Holland, 1978.
References 377
42. Clement, P.: Approximation by finite element functions using local regulariza-
tion. RAIRO Anal. Numer. R–2 (1975) 77–84.
43. Costabel, M.: Symmetric methods for the coupling of finite elements and bound-
ary elements. In: Boundary Elements IX (C. A. Brebbia, G. Kuhn, W. L. Wend-
land eds.), Springer, Berlin, pp. 411–420, 1987.
44. Costabel, M.: Boundary integral operators on Lipschitz domains: Elementary
results. SIAM J. Math. Anal. 19 (1988) 613–626.
45. Costabel, M., Stephan, E. P.: Boundary integral equations for mixed boundary
value problems in polygonal domains and Galerkin approximations. In: Math-
ematical Models and Methods in Mechanics. Banach Centre Publ. 15, PWN,
Warschau, pp. 175–251, 1985.
46. Dahmen, W., Prössdorf, S., Schneider, R.: Wavelet approximation methods for
pseudodifferential equations I: Stability and convergence. Math. Z. 215 (1994)
583–620.
47. Dahmen, W., Prössdorf, S., Schneider, R.: Wavelet approximation methods
for pseudodifferential equations II: Matrix compression and fast solution. Adv.
Comput. Math. 1 (1993) 259–335.
48. Darve, E.: The Fast Multipole Method: Numerical implementation. J. Comp.
Phys. 160 (2000) 195–240.
49. Dauge, M.: Elliptic boundary value problems on corner domains. Smoothness
and asymptotics of solutions. Lecture Notes in Mathematics, vol. 1341, Springer,
Berlin, 1988.
50. Dautray, R., Lions, J. L.: Mathematical Analysis and Numerical Methods for
Science and Technology. Volume 4: Integral Equations and Numerical Methods.
Springer, Berlin, 1990.
51. Demkowicz, L.: Computing with hp Adaptive Finite Elements. Vol. 1. Chapman
& Hall/CRC, Boca Raton, 2007.
52. Douglas, C. C., Haase, G., Langer, U.: A Tutorial on Elliptic PDE Solvers and
their Parallelization. SIAM, Philadelphia, 2003.
53. Duvaut, G., Lions, J. L.: Inequalities in Mechanics and Physics. Springer, Berlin,
1976.
54. Engleder, S., Steinbach, O.: Modified Boundary Integral Formulations for the
Helmholtz Equation. J. Math. Anal. Appl. 331 (2007) 396–407.
55. Erichsen, S., Sauter, S. A.: Efficient automatic quadrature in 3 − d Galerkin
BEM. Comp. Meth. Appl. Mech. Eng. 157 (1998) 215–224.
56. Faermann, B.: Local a posteriori error indicators for the Galerkin discretization
of boundary integral equations. Numer. Math. 79 (1998) 43–76.
57. Fix, G. J., Strang, G.: An Analysis of the Finite Element Method. Prentice Hall
Inc., Englewood Cliffs, 1973.
58. Fortin, M.: An analysis of the convergence of mixed finite element methods.
R.A.I.R.O. Anal. Numer. 11 (1977) 341–354.
59. Fox, L., Huskey, H. D., Wilkinson, J. H.: Notes on the solution of algebraic
linear simultaneous equations. Quart. J. Mech. Appl. Math. 1 (1948) 149–173.
60. Gatica, G. N., Hsiao, G. C.: Boundary–Field Equation Methods for a Class of
Nonlinear Problems. Pitman Research Notes in Mathematics Series, vol. 331.
Longman, Harlow, 1995.
61. Gaul, L., Kögl, M., Wagner, M.: Boundary Element Methods for Engineers and
Scientists. Springer, Berlin, 2003.
378 References
62. Giebermann, K.: Schnelle Summationsverfahren zur numerischen Lösung von In-
tegralgleichungen für Streuprobleme im R3 . Dissertation, Universität Karlsruhe,
1997.
63. Gradshteyn, I. S., Ryzhik, I. M.: Table of Integrals, Series, and Products. Aca-
demic Press, New York, 1980.
64. Greengard, L.: The Rapid Evaluation of Potential Fields in Particle Systems.
The MIT Press, Cambridge, MA, 1987.
65. Greengard, L., Rokhlin, V.: A fast algorithm for particle simulations. J. Comput.
Phys. 73 (1987) 325–348.
66. Grisvard, P.: Elliptic Problems in Nonsmooth Domains. Pitman, Boston, 1985.
67. Guiggiani, G., Gigante, A.: A general algorithm for multidimensional Cauchy
principal value integrals in the boundary element method. ASME J. Appl. Mech.
57 (1990) 906–915.
68. Haase, G.: Parallelisierung numerischer Algorithmen für partielle Differential-
gleichungen. B. G. Teubner, Stuttgart, Leipzig, 1999.
69. Hackbusch, W.: Multi–Grid Methods and Applications. Springer, Berlin, 1985.
70. Hackbusch, W.: Iterative Lösung grosser schwachbesetzter Gleichungssysteme.
B. G. Teubner, Stuttgart, 1993.
71. Hackbusch, W.: Theorie und Numerik elliptischer Differentialgleichungen.
B. G. Teubner, Stuttgart, 1996.
72. Hackbusch, W.: A sparse matrix arithmetic based on H–matrices. I. Introduc-
tion to H–matrices. Computing 62 (1999) 89–108.
73. Hackbusch, W., Nowak, Z. P.: On the fast matrix multiplication in the boundary
element method by panel clustering. Numer. Math. 54, 463–491 (1989).
74. Hackbusch, W., Wittum, G. (eds.): Boundary Elements: Implementation and
Analysis of Advanced Algorithms. Notes on Numerical Fluid Mechanics 54,
Vieweg, Braunschweig, 1996.
75. Han, H.: The boundary integro–differential equations of three–dimensional Neu-
mann problem in linear elasticity. Numer. Math. 68 (1994) 269–281.
76. Harbrecht, H.: Wavelet Galerkin schemes for the boundary element method in
three dimensions. Doctoral Thesis, TU Chemnitz, 2001.
77. Hartmann, F.: Introduction to Boundary Elements. Springer, Berlin, 1989.
78. Hestenes, M., Stiefel, E.: Methods of conjugate gradients for solving linear sys-
tems. J. Res. Nat. Bur. Stand 49 (1952) 409–436.
79. Hörmander, L.: The Analysis of Linear Partial Differential Operators I, Springer,
Berlin, 1983.
80. Hsiao, G. C., Stephan, E. P., Wendland, W. L.: On the integral equation method
for the plane mixed boundary value problem of the Laplacian. Math. Meth.
Appl. Sci. 1 (1979) 265–321.
81. Hsiao, G. C., Wendland, W. L.: A finite element method for some integral
equations of the first kind. J. Math. Anal. Appl. 58 (1977) 449–481.
82. Hsiao, G. C., Wendland, W. L.: The Aubin–Nitsche lemma for integral equa-
tions. J. Int. Equat. 3 (1981) 299–315.
83. Hsiao, G. C., Wendland, W. L.: Integral Equation Methods for Boundary Value
Problems. Springer, Heidelberg, to appear.
84. Jaswon, M. A., Symm, G. T.: Integral Equation Methods in Potential Theory
and Elastostatics. Academic Press, London, 1977.
85. Jung, M., Langer, U.: Methode der finiten Elemente für Ingenieure. B. G.
Teubner, Stuttgart, Leipzig, Wiesbaden, 2001.
References 379
86. Jung, M., Steinbach, O.: A finite element–boundary element algorithm for in-
homogeneous boundary value problems. Computing 68 (2002) 1–17.
87. Kieser, R., Schwab, C., Wendland, W. L.: Numerical evaluation of singular and
finite–part integrals on curved surfaces using symbolic manipulation. Computing
49 (1992) 279–301.
88. Kress, R.: Linear Integral Equations. Springer, Heidelberg, 1999.
89. Kupradze, V. D.: Three–dimensional problems of the mathematical theory of
elasticity and thermoelasticity. North–Holland, Amsterdam, 1979.
90. Kythe, P. K.: Fundamental Solutions for Differential Operators and Applica-
tions. Birkhäuser, Boston, 1996.
91. Ladyzenskaja, O. A.: Funktionalanalytische Untersuchungen der Navier–
Stokesschen Gleichungen. Akademie–Verlag, Berlin, 1965.
92. Ladyzenskaja, O. A., Ural’ceva, N. N.: Linear and quasilinear elliptic equations.
Academic Press, New York, 1968.
93. Lage, C., Sauter, S. A.: Transformation of hypersingular integrals and black–box
cubature. Math. Comp. 70 (2001) 223–250.
94. Lage, C., Schwab, C.: Wavelet Galerkin algorithms for boundary integral equa-
tions. SIAM J. Sci. Comput. 20 (1999) 2195–2222.
95. Langer, U.: Parallel iterative solution of symmetric coupled fe/be equations via
domain decomposition. Contemp. Math. 157 (1994) 335–344.
96. Langer, U., Pusch, D.: Data–sparse algebraic multigrid methods for large scale
boundary element equations. Appl. Numer. Math. 54 (2005) 406–424.
97. Langer, U., Steinbach, O.: Coupled boundary and finite element tearing and
interconnecting methods. In: Domain Decomposition Methods in Science and
Engineering (R. Kornhuber et. al. eds.), Lecture Notes in Computational Science
and Engineering, vol. 40, Springer, Heidelberg, pp. 83–97, 2004.
98. Langer, U., Steinbach, O., Wendland, W. L.: Computing and Visualization in
Science. Special Issue on Fast Boundary Element Methods in Industrial Appli-
cations. Volume 8, Numbers 3–4, 2005.
99. Maischak, M.: hp–Methoden für Randintegralgleichungen bei 3D–Problemen.
Theorie und Implementierung. Doctoral Thesis, Universität Hannover, 1995.
100. Maischak, M., Stephan, E. P.: The hp–version of the boundary element method
in R3 . The basic approximation results. Math. Meth. Appl. Sci. 20 (1997) 461–
476.
101. Maue, A. W.: Zur Formulierung eines allgemeinen Beugungsproblems durch
eine Integralgleichung. Z. f. Physik 126 (1949) 601–618.
102. Mazya, V. G.: Boundary integral equations. In: Analysis IV (V. G. Mazya,
S. M. Nikolskii eds.), Encyclopaedia of Mathematical Sciences, vol. 27, Springer,
Heidelberg, pp. 127–233, 1991.
103. McLean, W.: Strongly Elliptic Systems and Boundary Integral Equations.
Cambridge University Press, 2000.
104. McLean, W., Steinbach, O.: Boundary element preconditioners for a hyper-
singular boundary integral equation on an intervall. Adv. Comput. Math. 11
(1999) 271–286.
105. McLean, W., Tran, T.: A preconditioning strategy for boundary element
Galerkin methods. Numer. Meth. Part. Diff. Eq. 13 (1997) 283–301.
106. Nec̆as, J.: Les Methodes Directes en Theorie des Equations Elliptiques. Masson,
Paris und Academia, Prag, 1967.
107. Nedelec, J. C.: Integral equations with non integrable kernels. Int. Eq. Operator
Th. 5 (1982) 562–572.
380 References
108. Of, G., Steinbach, O.: A fast multipole boundary element method for a modified
hypersingular boundary integral equation. In: Analysis and Simulation of Mul-
tifield Problems (W. L. Wendland, M. Efendiev eds.), Lecture Notes in Applied
and Computational Mechanics 12, Springer, Heidelberg, 2003, pp. 163–169.
109. Of, G., Steinbach, O., Wendland, W. L.: Applications of a fast multipole
Galerkin boundary element method in linear elastostatics. Comput. Vis. Sci.
8 (2005) 201–209.
110. Of, G., Steinbach, O., Wendland, W. L.: The fast multipole method for the
symmetric boundary integral formulation. IMA J. Numer. Anal. 26 (2006) 272–
296.
111. Petersdorff, T. von, Stephan, E. P.: On the convergence of the multigrid method
for a hypersingular integral equation of the first kind. Numer. Math. 57 (1990)
379–391.
112. Plemelj, J.: Potentialtheoretische Untersuchungen. Teubner, Leipzig, 1911.
113. Prössdorf, S., Silbermann, B.: Numerical Analysis for Integral and Related
Operator Equations. Birkhäuser, Basel, 1991.
114. Quarteroni, A., Valli, A.: Domain Decomposition Methods for Partial Differ-
ential Equations. Oxford Science Publications, 1999.
115. Rathsfeld, A.: Quadrature methods for 2D and 3D problems. J. Comput. Appl.
Math. 125 (2000) 439–460.
116. Reidinger, B., Steinbach, O.: A symmetric boundary element method for the
Stokes problem in multiple connected domains. Math. Meth. Appl. Sci. 26 (2003)
77–93.
117. Rjasanow, S., Steinbach, O.: The Fast Solution of Boundary Integral Equa-
tions. Mathematical and Analytical Techniques with Applications to Engineer-
ing. Springer, New York, 2007.
118. Rudin, W.: Functional Analysis. McGraw–Hill, New York, 1973.
119. Ruotsalainen, K., Wendland, W. L.: On the boundary element method for some
nonlinear boundary value problems. Numer. Math. 53 (1988) 299–314.
120. Saad, Y., Schultz, M. H.: A generalized minimal residual algorithm for solving
nonsymmetric linear systems. SIAM J. Sci. Stat. Comput. 7 (1985) 856–869.
121. Sauer, T., Xu, Y.: On multivariate Lagrange interpolation. Math. Comp. 64
(1995) 1147–1170.
122. Sauter, S. A.: Variable order panel clustering. Computing 64 (2000) 223–261.
123. Sauter, S. A., Schwab, C.: Quadrature of hp–Galerkin BEM in R3 . Numer.
Math. 78 (1997) 211–258.
124. Sauter, S. A., Schwab, C.: Randelementmethoden. Analyse, Numerik und
Implementierung schneller Algorithmen. B. G. Teubner, Stuttgart, Leipzig,
Wiesbaden, 2004.
125. Schanz, M., Steinbach, O. (eds.): Boundary Element Analysis: Mathematical
Aspects and Applications. Lecture Notes in Applied and Computational Me-
chanics, vol. 29, Springer, Heidelberg, 2007.
126. Schatz, A. H., Thomée, V., Wendland, W. L.: Mathematical Theory of Finite
and Boundary Element Methods. Birkhäuser, Basel, 1990.
127. Schneider, R.: Multiskalen– und Wavelet–Matrixkompression: Analysisbasierte
Methoden zur effizienten Lösung grosser vollbesetzter Gleichungssysteme. Ad-
vances in Numerical Mathematics. B. G. Teubner, Stuttgart, 1998.
128. Schulz, H., Steinbach, O.: A new a posteriori error estimator in adaptive direct
boundary element methods. Calcolo 37 (2000) 79–96.
References 381
129. Schulz, H., Wendland, W. L.: Local a posteriori error estimates for boundary
element methods. In: ENUMATH 2007, World Sci. Publ., River edge, pp. 564–
571, 1998.
130. Schwab, C.: p– and hp–Finite Element Methods. Theory and Applications in
Solid and Fluid Mechanics. Clarendon Press, Oxford, 1998.
131. Schwab, C., Suri, M.: The optimal p–version approximation of singularities on
polyhedra in the boundary element method. SIAM J. Numer. Anal. 33 (1996)
729–759.
132. Schwab, C., Wendland, W. L.: On numerical cubatures of singular surface
integrals in boundary element methods. Numer. Math. 62 (1992) 343–369.
133. Scott, L. R., Zhang, S.: Finite element interpolation of nonsmooth functions
satisfying boundary conditions. Math. Comp. 54 (1990) 483–493.
134. Sirtori, S.: General stress analysis method by means of integral equations and
boundary elements. Meccanica 14 (1979) 210–218.
135. Sloan, I. H.: Error analysis of boundary integral methods. Acta Numerica 92
(1992) 287–339.
136. Steinbach, O.: Fast evaluation of Newton potentials in boundary element meth-
ods. East–West J. Numer. Math. 7 (1999) 211–222.
137. Steinbach, O.: On the stability of the L2 projection in fractional Sobolev spaces.
Numer. Math. 88 (2001) 367–379.
138. Steinbach, O.: On a generalized L2 projection and some related stability esti-
mates in Sobolev spaces. Numer. Math. 90 (2002) 775–786.
139. Steinbach, O.: Stability estimates for hybrid coupled domain decomposition
methods. Lecture Notes in Mathematics 1809, Springer, Heidelberg, 2003.
140. Steinbach, O.: Numerische Näherungsverfahren für elliptische Randwertprob-
leme. Finite Elemente und Randelemente. B. G. Teubner, Stuttgart, Leipzig,
Wiesbaden, 2003.
141. Steinbach, O.: A robust boundary element method for nearly incompressible
elasticity. Numer. Math. 95 (2003) 553–562.
142. Steinbach, O.: A note on the ellipticity of the single layer potential in two–
dimensional linear elastostatics. J. Math. Anal. Appl. 294 (2004) 1–6.
143. Steinbach, O.: Lösungsverfahren für lineare Gleichungssysteme. Algorithmen
und Anwendungen. B. G. Teubner, Stuttgart, Leipzig, Wiesbaden, 2005.
144. Steinbach, O., Wendland, W L.: The construction of some efficient precondi-
tioners in the boundary element method. Adv. Comput. Math. 9 (1998) 191–216.
145. Steinbach, O., Wendland, W. L.: On C. Neumann’s method for second order
elliptic systems in domains with non–smooth boundaries. J. Math. Anal. Appl.
262 (2001) 733–748.
146. Stephan, E. P.: The h–p version of the boundary element method for solving
2– and 3–dimensional problems. Comp. Meth. Appl. Mech. Eng. 133 (1996)
183–208.
147. Stephan, E. P.: Multilevel methods for the h–, p–, and hp–versions of the
boundary element method. J. Comput. Appl. Math. 125 (2000) 503–519.
148. Stroud, A. H.: Approximate Calculations of Multiple Integrals. Prentice Hall,
Englewood Cliffs, 1973.
149. Toselli, A., Widlund, O.: Domain Decomposition Methods–Algorithms and
Theory. Springer, Berlin, 2005.
150. Triebel, H.: Höhere Analysis. Verlag Harri Deutsch, Frankfurt/M., 1980.
151. Tyrtyshnikov, E. E.: Mosaic–skeleton approximations. Calcolo 33 (1996) 47–57.
382 References
152. Verchota, G.: Layer potentials and regularity for the Dirichlet problem for
Laplace’s equation in Lipschitz domains. J. Funct. Anal. 59 (1984) 572–611.
153. Verfürth, R.: A Review of A Posteriori Error Estimation and Adaptive Mesh–
Refinement. John Wiley & Sons, Chichester, 1996.
154. Vladimirov, V. S.: Equations of Mathematical Physics. Marcel Dekker, New
York, 1971.
155. van der Vorst, H. A.: Bi–CGSTAB: A fast and smoothly converging variant
of Bi–CG for the solution of nonsymmetric linear systems. SIAM J. Sci. Stat.
Comput. 13 (1992) 631–644.
156. Walter, W.: Einführung in die Theorie der Distributionen. BI Wissenschaftsver-
lag, Mannheim, 1994.
157. Wendland, W. L.: Elliptic Systems in the Plane. Pitman, London, 1979.
158. Wendland, W. L. (ed.): Boundary Element Topics. Springer, Heidelberg, 1997.
159. Wendland, W. L., Zhu, J.: The boundary element method for three–
dimensional Stokes flow exterior to an open surface. Mathematical and Com-
puter Modelling 15 (1991) 19–42.
160. Wloka, J.: Funktionalanalysis und Anwendungen. Walter de Gruyter, Berlin,
1971.
161. Wloka, J.: Partielle Differentialgleichungen. B. G. Teubner, Stuttgart, 1982.
162. Xu, J.: An introduction to multilevel methods. In: Wavelets, multilevel methods
and elliptic PDEs. Numer. Math. Sci. Comput., Oxford University Press, New
York, 1997, pp. 213–302.
163. Yosida, K.: Functional Analysis. Springer, Berlin, Heidelberg, 1980.
Index
jump relation
edge, 203
adjoint double layer potential, 123
elliptic operator, 1, 46
double layer potential, 127
single layer potential, 120
far field, 331
finite element, 203 Korn’s inequality
form function first, 74
Bubble–, 215 second, 76
constant, 211
linear, 212 Lagrange
quadratic, 214 functional, 53
Fourier transform, 30 multiplier, 52, 64
Fredholm alternative, 58 multipliers, 255
fundamental solution, 90 Lamé constants, 6
Helmholtz operator, 106, 109 Lax–Milgram lemma, 46
Laplace operator, 96 Lipschitz domain, 20
linear elasticity, 100 load vector, 268
Stokes system, 103 locally quasi–uniform, 205